Skip to main content


See other formats



tie urn if mm uism smisnts 


"The narrative flows easily, and all the points are driven home with 
engaging examples from real life. I found Best's book a delight. 
Always engaging, it is accessible to a lay reader, yet will reward the 
expert; the examples it gives could enrich both a primary schoolroom 
and a university lecture hall." 

N a tur e 

"Invaluable counsel for good citizenship." 

— Booklist 

"This informative and well- written little book will be a particularly 
worthwhile addition to libraries' collections and will help all readers 
become savvier and more critical news consumers.' 

— Publishers Weekly 

"Whether we like them or not, we have to live with statistics, and 
Damned Lies and Statistics offers a useful guide for engaging with 
their troublesome world. Despite the temptation to be cynical, the au- 
thor of this timely and excellent work cautions the reader against re- 
acting in such a way to statistics. What we are offered is an approach 
that helps us to work out the real story behind those numbers." 

The Independent 

"Deserves a place next to the dictionary on every school, media, and 
home-office desk. " 

The Boston Globe 

"A clearly written primer for the statistically impaired. It is as impor- 
tant to discussions of public policy as any book circulating today." 

— The Christian Science Monitor 

"Definitely a must for politicians, activists and others who generate or 
use statistics, but especially for those who want to think for them- 
selves rather than take as gospel every statistic presented to them." 

— New Scientist 

"Damned Lies and Statistics is highly entertaining as well as instructive. 
Best's book shows how some of those big numbers indicating big so- 
cial problems were created in the first place and instructs the reader 
(and reporter) how to be on guard against such gross manipulation. 
And it doesn't take an understanding of advanced mathematics to do 
so thanks to this book, which ought to be required reading in every 
newsroom in the country." 

— The Washington Times 





University of California Press 
Berkeley and Los Angeles, California 

University of California Press, Ltd. 

London, England 

© 2004 by the Regents of the University of California 
Library of Congress Cataloging -in-Publication Data 
Best, Joel. 

More damned lies and statistics | how numbers confuse 
public issues / Joel Best, 
p. cm. 

Includes bibliographical references and index. 

ISBN 0-520- 333 

1. Sociology — Statistical methods. 2. Social problems- 
Statisti cal methods. 3. Social indicators. L Title. 

HM535.B474 2004 

303.3 '8 — dc22 2003028076 

Manufactured in the United States of America 

13 12 11 10 09 08 07 06 05 04 

10 987654 321 

Printed on Ecobook 50 containing a minimum 50% post- 
consumer waste, processed chlorine free. The balance contains 
virgin pulp, including 25% Forest Stewardship Council Certified 
for no old growth tree cutting, processed either tof or eof. The 
sheet is acid-free and meets the minimum requirements of 
ansi/niso 239.48-1992(1* 1997) (Permanence of Paper). 


Acknowledgments 1X 

Preface: People Count X1 

1. Missing Numbers 1 

2. Confusing Numbers 2 6 

3. Scary Numbers 63 

4. Authoritative Numbers 9 1 

5. Magical Numbers 1 

6. Contentious Numbers : 44 

7. Toward Statistical Literacy? I 7° 

Notes 1 83 

Index x 97 


I had not planned to write this book. It is a sequel to my 
Damned Lies and Statistics (DLS), which was published in 
2001. When I finished writing DLS, I thought that I was 
through writing about statistics, and I had plans to begin 
working on a completely different project. Besides, I'm a pro- 
fessor, and professors don't get opportunities to write sequels — 
we feel fortunate if somebody is willing to publish, let alone 
read, what we write even once. 

However, almost as soon as DLS appeared, I began getting 
e-mail messages from people who had read the book. Often, they 
drew my attention to wonderfully dubious statistics reported in 
the media. Among my favorites: a newspaper columnist who 
warned that smoking "kills one in five Americans each year"; 
and a British news item suggesting that "40 percent of young 
men have such a poor grasp of the way a bra fastens that they 
risk serious finger injuries." Others wrote to suggest topics that 

DLS hadn't treated (some messages were from college instruc- 
tors frustrated by the difficulties of conveying particular points 
in their courses). 

I also began receiving invitations to talk to groups or write 
about statistics; often, I was asked to address particular topics 
that were not familiar to me. Studying new subjects sometimes 
raised new issues that I began to wish I'd addressed in DLS. 

So when Naomi Schneider, my editor at the University of 
California Press, asked whether I might like to write a sequel to 
DLS , I agreed. I'd begun believing that I had enough ideas for 
another book, and there seemed to be enough people interested 
in the topic. I'm afraid I've lost track of the sources for some of 
my ideas, but I can at least thank those folks who I know made 
suggestions that were, in one way or another, incorporated in 
this book, along with thanking those who read and commented 
on parts of the manuscript. These include, in addition to 
Naomi, David Altheide, Ronet Bachman, Joan Best, George 
Bizer, Barbara Costello, Michael Gallagher, Linda Gottfredson, 
Larry Griffith, Henry Hipkens, Jim Holstein, Philip Jenkins, 
Vivian Klaff, the late Carl Klockars, Kathe Lowney, Katherine 
C. MacKinnon, Michael J. McFadden, Eric Rise, Naomi B. 
Robbins, Milo Schield, and — I fear — others whose names were 
inadvertently misplaced. I especially want to thank Vicky 
Baynes for helping me with the mysterious process of turning 
graphs into computer files. These people, of course, should be 
credited for providing help but not blamed for my interpreta- 
tions. Thank you all. I hope this new book pleases you. 


L unch was at a prominent conservative think tank. The 
people around the table were fairly well known; I’d read 
some of their books and articles and had even seen them 
interviewed on television. They listened to me talk 
about bad statistics, and they agreed that the problem was seri- 
ous. They had only one major criticism: I’d missed the role of 
ideology. Bad statistics, they assured me, were almost always 
promoted by liberals. 

Two months earlier, I'd been interviewed by a liberal radio 
talk-show host (they do exist!). He, too, thought it was high 
time to expose bad statistics — especially those so often circulat- 
ed by conservatives. 

When I talk to people about statistics, I find that they usually 
are quite willing to criticize dubious statistics — as long as the 
numbers come from people with whom they disagree. Political 
conservatives are convinced that the statistics presented by lib- 

erals are deeply flawed, just as liberals are eager to denounce 
conservatives' shaky figures. When conservatives (or liberals) 
ask me how to spot bad statistics, I suspect that they'd like me 
to say, "Watch out for numbers promoted by people with whom 
you disagree." Everyone seems to insist that the other guy's 
figures are lousy (but mine are, of course, just fine, or at least 
good enough). People like examples of an opponent's bad statis- 
tics, but they don't care to have their own numbers criticized be- 
cause, they worry, people might get the wrong idea: criticizing 
my statistics might lead someone to question my larger argu- 
ment, so let's focus on the other guy's errors and downplay 

Alas, I don't believe that any particular group, faction, or ide- 
ology holds a monopoly on poor statistical reasoning. In fact, in 
choosing examples to illustrate this book's chapters, I've tried to 
identify a broad range of offenders. My goal is not to convince 
you that those other guys can't be trusted (after all, you proba- 
bly already believe that). Rather, I want you to come away from 
this book with a sense that all numbers — theirs and yours — 
need to be handled with care. 

This is tricky, because we tend to assume that statistics are 
facts, little nuggets of truth that we uncover, much as rock col- 
lectors find stones.' After all, we think, a statistic is a number, 
and numbers seem to be solid, factual proof that someone must 
have actually counted something. But that's the point: people 
count. For every number we encounter, some person had to do 
the counting. Instead of imagining that statistics are like rocks, 
we'd do better to think of them as jewels. Gemstones may be 
found in nature, but people have to create jewels. Jewels must 
be selected, cut, polished, and placed in settings to be viewed 

from particular angles. In much the same way, people create sta- 
tistics: they choose what to count, how to go about counting, 
which of the resulting numbers they share with others, and 
which words they use to describe and interpret those figures. 
Numbers do not exist independent of people; understanding 
numbers requires knowing who counted what, why they both- 
ered counting, and how they went about it. 

All statistics are products of social activity, the process sociol- 
ogists call social construction. Although this point might seem 
painfully obvious, it tends to be forgotten or ignored when we 
think about — and particularly when we teach — statistics. We 
usually envision statistics as a branch of mathematics, a view re- 
inforced by high school and college statistics courses, which 
begin by introducing probability theory as a foundation for sta- 
tistical thinking, a foundation on which is assembled a structure 
of increasingly sophisticated statistical measures. Students are 
taught the underlying logic of each measure, the formula used 
to compute the measure, the software commands that can ex- 
tract it from the computer, and some guidelines for interpreting 
the numbers that result from these computations. These are 
complicated lessons: few students have an intuitive grasp of any 
but the simplest statistics, and instruction usually focuses on 
clarifying the computational complexities. 

The result is that statistical instruction tends to downplay 
consideration of how real-life statistics come into being. Yet all 
statistics are products of people's choices and compromises, 
which inevitably shape, limit, and distort the outcome. Statistics 
instructors often dismiss this as melodramatic irrelevance. Just 
as the conservatives at the think tank lunch imagined that bad 
statistics were the work of devious liberals, statistics instructors 

might briefly caution that calculations or presentations of statis- 
tical results may be "biased" (that is, intentionally designed to 
deceive). Similarly, a surprisingly large number of book titles 
draw a distinction between statistics and lies: How to Lie with 
Statistics (also, How to Lie with Charts, How to Lie with Maps, 
and so on); How to Tell the Liars from the Statisticians; How 
Numbers Lie; even (ahem) my own Damned Lies and Statistics 
One might conclude that statistics are pure, unless they unfor- 
tunately become contaminated by the bad motives of dishonest 

Perhaps it is necessary to set aside the real world in an effort 
to teach students about advanced statistical reasoning. But dis- 
missive warnings to watch out for bias don't go very far in 
preparing people to think critically about the numbers they read 
in newspaper stories or hear from television commentators. 
Statistics play important roles in real-world debates about social 
problems and social policies; numbers become key bits of evi- 
dence used to challenge opponents' claims and to promote one's 
own views. Because people do knowingly present distorted or 
even false figures, we cannot dismiss bias as nonexistent. But 
neither can we simply categorize numbers as either true figures 
presented by sincere, well-meaning people (who, naturally, 
agree with us) or false statistics knowingly promoted by devious 
folks (who are on the other side, of course). 

Misplaced enthusiasm is probably at least as common as de- 
liberate bias in explaining why people spread bad statistics. 
Numbers rarely come first. People do not begin by carefully cre- 
ating some bit of statistical information and then deduce what 
they ought to think. Much more often, they start with their own 
interests or concerns, which lead them to run across, or perhaps 

actively uncover, relevant statistical information. When these 
figures support what people already believe — or hope, or fear — 
to be true, it is very easy for them to adopt the numbers, to over- 
look or minimize their limitations, to find the figures first ar- 
resting, then compelling, and finally authoritative. People soon 
begin sharing these now important numbers with others and be- 
come outraged if their statistics are questioned. One need not in- 
tentionally lie to others, or even to oneself. One need only let 
down one's critical guard when encountering a number that 
seems appealing, and momentum can do the rest. 

The solution is to maintain critical standards when thinking 
about statistics. Some people are adept at this, as long as they are 
examining their opponents' figures. It is much more difficult to 
maintain a critical stance toward our own numbers. After all, 
our numbers support what we believe to be true. Whatever 
minor flaws they might have surely must be unimportant. At 
least, that's what we tell ourselves when we justify having a 
double standard for judging our own statistics and those of 

This book promotes what we might call a single standard for 
statistical criticism. It argues that we must recognize that all 
numbers are social products and that we cannot understand a 
statistic unless we know something about the process by which 
it came into being. It further argues that all statistics are imper- 
fect and that we need to recognize and acknowledge their flaws 
and limitations. All this is true regardless of whether we agree 
or disagree with the people presenting the numbers. We need to 
think critically about both the other guys' figures and our own. 

I should confess that, in writing this book, I have done little 
original research. I have borrowed most of my examples from 

works by other analysts, mostly social scientists and journalists. 
My goal in writing about bad statistics is to show how these 
numbers emerge and spread. Just as I do not believe that this is 
the work of one political faction, I do not mean to suggest that 
all the blame can be laid at the door of one segment of society, 
such as the media. The media often circulate bad numbers, but 
then so do activists, corporations, officials, and even scientists — 
in fact, those folks usually are the sources for the statistics that 
appear in the media. And, we should remember, the problems 
with bad statistics often come to light through the critical efforts 
of probing journalists or scientists who think the numbers 
through, discover their flaws, and bring those flaws to public at- 
tention. A glance at my sources will reveal that critical thinking, 
just like bad statistics, can be found in many places. 

The chapters in this book explore some common problems in 
thinking about social statistics. The chapter titles refer to differ- 
ent sorts of numbers — missing numbers, confusing numbers, 
and so on. As I use them, these terms have no formal mathe- 
matical meanings; they are simply headings for organizing the 
discussion. Thus, chapter i addresses what I call missing num- 
bers, that is, statistics that might be relevant to debates over so- 
cial issues but that somehow don't emerge during those discus- 
sions. It identifies several types of missing numbers and seeks to 
account for their absence. Chapter 2 considers confusing num- 
bers, basic problems that bedevil our understanding of many 
simple statistics and graphs. Scary numbers — statistics about 
risks and other threats — are the focus of chapter 3. 

The next three chapters explore the relationship between au- 
thority and statistics. Chapter 4's subject is authoritative num- 
bers. This chapter considers what we might think of as statistics 

that seem good enough to be beyond dispute — products of 
scientific research or government data collection, for instance. It 
argues that even the best statistics need to be handled with care, 
that even data gathered by experts can be subject to misinter- 
pretation. Chapter 5 examines what I call magical numbers — 
efforts to resolve issues through statistics, as though figures are 
a way to distill reality into pure, incontrovertible facts. Chapter 
6 concentrates on contentious numbers, cases of data duels and 
stat wars in which opponents hurl contradictory figures at one 
another. Finally, chapter 7 explores the prospects for teaching 
statistical literacy, for improving public understanding of num- 
bers and teaching people how to be more thoughtful and more 
critical consumers of statistics. 

The lesson that people count — that we don't just find statis- 
tics but that we create them — offers both a warning and a 
promise. The warning is that we must be wary, that unless we 
approach statistics with a critical attitude, we run the risk of 
badly misunderstanding the world around us. But there is also 
a promise: that we need not be at the mercy of numbers, that we 
can learn to think critically about them, and that we can come 
to appreciate both their strengths and their flaws. 


C BS News anchor Dan Rather began his evening news- 
cast on March 5, 2001, by declaring: "School shootings 
in this country have become an epidemic." That day, a 
student in Santee, California, had killed two other stu- 
dents and wounded thirteen more, and media coverage linked 
this episode to a disturbing trend. Between December 1997 and 
May 1998, there had been three heavily publicized school shoot- 
ing incidents: in West Paducah, Kentucky (three dead, five 
wounded); Jonesboro, Arkansas (five dead, ten wounded); and 
Springfield, Oregon (two dead and twenty-one wounded at the 
school, after the shooter had killed his parents at home). The fol- 
lowing spring brought the rampage at Columbine High School 
in Littleton, Colorado, in which two students killed twelve fel- 
low students and a teacher, before shooting themselves.' Who 
could doubt Rather's claim about an epidemic? 

And yet the word epzdemzc suggests a widespread, growing 

phenomenon. Were school shootings indeed on the rise? Sur- 
prisingly, a great deal of evidence indicated that they were not: 

Since school shootings are violent crimes, we might begin 
by examining trends in criminality documented by the Federal 
Bureau of Investigation. The Uniform Crime Reports, the FBI's 
tally of crimes reported to the police, showed that the overall 
crime rate, as well as the rates for such major violent crimes as 
homicide, robbery, and aggravated assault, fell during the 1990s. 

. Similarly, the National Crime Victimization Survey (which 
asks respondents whether anyone in their household has been a 
crime victim) revealed that victimization rates fell during the 
1990s; in particular, reports of teenagers being victimized by vio- 
lent crimes at school dropped. 

Other indicators of school violence also showed decreases. 
The Youth Risk Behavior Survey conducted by the U.S. Centers 
for Disease Control and Prevention found steadily declining 
percentages of high school students who reported fighting or 
carrying weapons on school property during the 1990s. 

Finally, when researchers at the National School Safety 
Center combed media reports from the school years 1992-1993 
through 2000-2001, they identified 321 violent deaths that had 
occurred at schools. Not all of these incidents involved student- 
on-student violence; they included, for example, 16 accidental 
deaths and 56 suicides, as well as incidents involving nonstu- 
dents, such as a teacher killed by her estranged husband (who 
then shot himself) and a nonstudent killed on a school play- 
ground during a weekend. Even if we include all 32 1 of these 
deaths, however, the average fell from 48 violent deaths per year 
during the school years 1992—1993 through 1996—199710 32 per 

year from 1997-1998 through 2000—2001, If we eliminate acci- 
dental deaths and suicides, the decline remains, with the average 
falling from 31 deaths per year in the earlier period to 24 per 
year in the later period (which included all of the heavily publi- 
cized incidents mentioned earlier). While violent deaths are 
tragedies, they are also rare. Tens of millions of children attend 
school; for every million students, fewer than one violent death 
per year occurs in school. 

In other words, a great deal of statistical evidence was available 
to challenge claims that the country was experiencing a sudden 
epidemic of school shootings. The FBI's Uniform Crime Reports 
and the National Crime Victimization Survey in particular are 
standard sources for reporters who examine crime trends; the 
media's failure to incorporate findings from these sources in 
their coverage of school shootings is striking." 

Although it might seem that statistics appear in every discus- 
sion of every social issue, in some cases — such as the media's cov- 
erage of school shootings — relevant, readily available statistics 
are ignored. We might think of these as missing numbers. This 
chapter examines several reasons for missing numbers, includ- 
ing overwhelming examples, incalculable concepts, uncounted 
phenomena, forgotten figures, and legendary numbers. It asks 
why potentially relevant statistics don't figure in certain public 
debates and tries to assess the consequences of their absence. 


Why are numbers missing from some debates over social prob- 
lems and social policies? One answer is that a powerful example 

can overwhelm discussion of an issue. The 1999 shootings at 
Columbine High School are a case in point. The high death toll 
ensured that Columbine would be a major news story. Moreover, 
the school's location in a suburb of a major city made it easy for 
reporters to reach the scene. As it took some hours to evacuate the 
students and secure the building, the press had time to arrive and 
capture dramatic video footage that could be replayed to illus- 
trate related stories in the weeks that followed. The juxtaposition 
of a terrible crime in a prosperous suburban community made 
the story especially frightening — if this school shooting could 
happen at Columbine, surely such crimes could happen any- 
where. In addition, the Columbine tragedy occurred in the era of 
competing twenty-four-hour cable news channels; their decisions 
to run live coverage of several funeral and memorial services and 
to devote broadcast time to extended discussions of the event and 
its implications helped to keep the story alive for weeks. 

For today's media, a dramatic event can become more than 
simply a news story in its own right; reporters have become at- 
tuned to searching for the larger significance of an event so that 
they can portray newsworthy incidents as instances of a wide- 
spread pattern or problem. Thus, Columbine, when coupled 
with the earlier, heavily publicized school shooting stories of 
1997—1998, came to exemplify the problem of school violence. 
And, commentators reasoned, if a larger problem existed, it 
must reflect underlying societal conditions; that is, school shoot- 
ings needed to be understood as a trend, wave, or epidemic with 
identifiable causes. Journalists have been identifying such crime 
waves since at least the nineteenth century — and, for nearly as 
long, criminologists have understood that crime waves are not 
so much patterns in criminal behavior as they are patterns in 

media coverage. All of the available statistical evidence suggest- 
ed that school violence had declined from the early 1990s to the 
late 1990s; there was no actual wave of school shootings. But the 
powerful images from Columbine made that evidence irrele- 
vant. One terrible example was "proof" that school shootings 
were epidemic. 

Compelling examples need not even be true. The stories that 
folklorists call contemporary legends (or the more familiar term 
urban legends) also shape our thinking about social problems. 
Contemporary legends usually spread through informal chan- 
nels, which once meant word of mouth but now also includes 
the more modern means of faxes and e-mail messages. A leg- 
end’s key quality remains unchanged, however: it must be a 
good story, good enough for people to remember it and want to 
pass it along. Legends thrive because they arouse fear, disgust, 
or other powerful emotions that make the tales memorable and 
repeatable? Very often, contemporary legends are topical: when 
child abductions are in the news, we tell stories about kidnap- 
pings in shopping malls; when gangs are receiving attention, we 
warn each other about lethal gang initiation rites. Such stories 
shape our thinking about social problems in much the same way 
dramatic news stories do. 

The power of examples is widely recognized. A reporter 
preparing a story about any broad social condition — say, home- 
lessness — is likely to begin by illustrating the problem with an 
example, perhaps a particular homeless person. Journalists (and 
their editors) prefer interesting, compelling examples that will 
intrigue their audience. And advocates who are trying to pro- 
mote particular social policies learn to help journalists by guid- 
ing them to examples that can be used to make specific points. 

Thus, activists calling for increased services for the homeless 
might showcase a homeless family, perhaps a mother of young 
children whose husband has been laid off by a factory closing 
and who cannot find affordable housing. In contrast, politicians 
seeking new powers to institutionalize the homeless mentally ill 
might point to a deranged, violent individual who seems to en- 
danger passersby.' The choice of examples conveys a sense of a 
social problem's nature. 

The problem with examples — whether they derive from 
dramatic events, contemporary legends, or the strategic choices 
of journalists or advocates — is that they probably aren't espe- 
cially typical. Examples compel when they have emotional 
power, when they frighten or disturb us. But atypical examples 
usually distort our understanding of a social problem; when we 
concentrate on the dramatic exception, we tend to overlook 
the more common, more typical — but more mundane — cases. 
Thus, Democrats used to complain about Republican President 
Ronald Reagan's fondness for repeating the story of a "welfare 
queen" who had supposedly collected dozens of welfare checks 
using false identities? Using such colorful examples to typify 
welfare fraud implies that welfare recipients are undeserving or 
don't really need public assistance. Defenders of welfare often 
countered Reagan's anecdotes with statistics showing that recip- 
ients were deserving (as evidenced by the small number of able- 
bodied adults without dependent children who received bene- 
fits) or that criminal convictions for fraud were relatively few , 6 
The danger is that the powerful but atypical example — the 
homeless intact family, the welfare queen — will warp our vi- 
sion of a social problem, thereby reducing a complicated social 
condition to a simple, melodramatic fable. 

Statistics, then, offer a way of checking our examples. If stud- 
ies of the homeless find few intact families (or individuals who 
pose threats of violence), or if studies of welfare recipients find 
that fraud involving multiple false identities is rare, then we 
should recognize the distorting effects of atypical examples and 
realize that the absence of numbers can damage our ability to 
grasp the actual dimensions of our problems. 


Sometimes numbers are missing because phenomena are very 
hard to count. Consider another crime wave. During the sum- 
mer of 2002, public concern turned to kidnapped children. 
Attention first focused on the case of an adolescent girl abduct- 
ed from her bedroom one night — a classic melodramatic exam- 
ple of a terrible crime that seemingly could happen to anyone. 
As weeks passed without a sign of the girl, both the search and 
the accompanying news coverage continued. Reports of other 
cases of kidnapped or murdered children began linking these 
presumably unrelated crimes to the earlier kidnapping, leading 
the media to begin talking about an epidemic of abductions. 

This issue had a history, however. Twenty years earlier, ac- 
tivists had aroused national concern about the problem of miss- 
ing children by coupling frightening examples to large statisti- 
cal estimates. One widespread claim alleged that nearly two 
million children went missing each year, including fifty thou- 
sand kidnapped by strangers. Later, journalists and social scien- 
tists exposed these early estimates as being unreasonably high. 
As a result, in 2002, some reporters questioned the claims of a 
new abduction epidemic; in fact, they argued, the FBI had in- 

vestigated more kidnappings the previous year, which suggest- 
ed that these crimes were actually becoming less common? 

Both sets of claims — that kidnappings were epidemic and that 
they were declining — were based on weak evidence. Missing- 
children statistics can never be precise because missing children 
are so difficult to count. We encounter problems of definition: 

What is a child — that is, what is the upper age limit for 
being counted? 

. What do we mean by missing? How long must a child be 
missing to be counted — a few minutes, one day, seventy-two 

What sorts of absences should be counted? Wandering off 
and getting lost? Running away? Being taken by a relative dur- 
ing a family dispute? Is a child who is with a noncustodial par- 
ent at a known location considered missing? 

People need to agree about what to count before they can start 
counting, but not everyone agrees about the answers to these 
questions. Obviously, the answers chosen will affect the num- 
bers counted; using a broad definition means that more missing 
children will be counted. 

A second set of problems concerns reporting. Parents of 
missing children presumably call their local law enforcement 
agency — usually a police or sheriffs department. But those au- 
thorities may respond in different ways. Some states require 
them to forward all missing-children reports to a statewide 
clearinghouse, which is supposed to contact all law enforcement 
agencies in the state in order to facilitate the search. The clear- 
inghouses — and some departments — may notify the National 

Crime Information Center, a branch of the FBI that compiles 
missing-persons reports. Some reports also reach the National 
Center for Missing and Exploited Children (the federally fund- 
ed group best known for circulating pictures of missing chil- 
dren) or FBI investigators (who claim jurisdiction over a few, 
but by no means most, kidnappings). Authorities in the same ju- 
risdiction do not necessarily handle all missing-children reports 
the same way; the case of a six-year-old seen being dragged into 
a strange car is likely to be treated differently than a report of a 
sixteen-year- old who has run away. We can suspect that the poli- 
cies of different agencies will vary significantly. The point is that 
the jurisdiction from which a child disappears and the particu- 
lars of the case probably affect whether a particular missing- 
child report finds its way into various agencies' records. 

It is thus very difficult to make convincing comparisons of 
the numbers of missing children from either time to time or 
place to place. Reporters who noted that fewer child-kidnap- 
ping reports were filed with the FBI in 2002 than in 2001, and 
who therefore concluded that the problem was declining, mis- 
takenly assumed that the FBI's records were more complete and 
authoritative than they actually were. Some things — like miss- 
ing children — are very difficult to count, which should make us 
skeptical about the accuracy of statistics that claim to describe 
the situation. 

Such difficulties can create special problems when people try 
to weigh things that are relatively easy to measure against things 
that are less calculable. Consider the method of cost-benefit 
analysis as a basis for decision-making. 8 In principle, it seems 
straightforward: calculate the expected costs and the value of 
the expected benefits for different courses of action, and choose 

the option that promises the best outcome. One problem, how- 
ever, is that some costs and benefits are easier to compute than 
others. A teenager trying to decide whether to go to a movie or 
spend an evening babysitting can probably assign reasonably ac- 
curate dollar values to these options — the cost of the movie 
ticket and refreshments versus the expected earnings from 
babysitting — but even then the decision will probably hinge on 
additional assumptions about happiness: would I be happier 
spending the evening with my friends at a movie, or would I 
prefer to earn money that can be spent for some greater benefit 
down the line? 

When applied to questions of social policy, such calculations 
only become more complex. Should we build more highways or 
support mass transit? Mass transit is rarely self-supporting: if 
the cost per trip seems too high, riders abandon mass transit; in 
order to keep them riding, ticket prices usually must be kept 
low by subsidizing the system. Critics of mass transit sometimes 
argue that such subsidies are wrong, that mass transit is in- 
efficient, expensive, and therefore not competitive. Advocates 
respond that this critique ignores many of the relevant costs and 
benefits. Whereas riders directly bear the costs of using mass 
transit each time they buy a ticket, the ways we pay for the costs 
of highway travel are less obvious (for example, through gaso- 
line taxes). Moreover, highways carry hidden, quality of life 
costs, such as greater air pollution, more traffic fatalities, and 
cities that discourage foot traffic by devoting huge areas to roads 
and parking lots. But such costs are hard to calculate. Even if we 
can agree on the likely health costs from air pollution and traffic 
accidents, how can we hope to assign a dollar value to being able 
to comfortably walk from one destination to another? And, of 

course, the critics have a rebuttal: costs are also incurred in 
building and maintaining mass transit systems. And what about 
the freedom cars offer — the ability to choose your own route 
and schedule? Shouldn't these considerations be incorporated 
in any calculations? 

There are basically two solutions to the problems that intan- 
gible factors pose to cost-benefit analyses, but neither solution is 
completely satisfactory. The first is to leave these factors out of 
the equation, to simply ignore what seems impossible to quan- 
tify. But should factors such as quality of life be treated as irrel- 
evant simply because they are hard to measure? The second so- 
lution is to estimate the values of costs and benefits, to assign 
dollar values to them. This approach keeps these factors in view, 
but the process is obviously arbitrary — what dollar value 
should be assigned to comfort or freedom? It is easy to skew the 
results of any cost-benefit analysis by pegging values as either 
very high or very low. 

Our culture has a particularly difficult time assigning values 
to certain types of factors. Periodically, for example, the press 
expresses shock that a cost-benefit analysis has assigned some 
specific value to individual lives? Such revelations produce pre- 
dictably outraged challenges: how can anyone place a dollar 
value on a human life — aren't people's lives priceless? The an- 
swer to that question depends on when and where it is asked. 
Americans' notion that human life is priceless has a surprising- 
ly short history. Only a century ago, the parents of a child killed 
by a streetcar could sue the streetcar company for damages 
equal to the child's economic value to the family (basically, the 
child's expected earnings until adulthood); today, of course, the 
parents would sue for the (vastly greater) value of their pain and 

suffering. Even the dollar value of a child's life varies across 
time and space. 10 

But the larger point is that trade-offs are inevitable. Building 
a bridge or implementing a childhood vaccination program has 
both risks and costs — as do the alternatives of not building the 
bridge or not vaccinating children. Our culture seems to have a 
lot of difficulty debating whether, say, vaccinations should pro- 
ceed if they will cause some number of children to sicken and 
die. Advocates on both sides try to circumvent this debate by 
creating melodramatically simple alternatives: vaccine propo- 
nents can be counted on to declare that harm from vaccines is 
virtually nonexistent but that failure to vaccinate will have ter- 
rible, widespread consequences; whereas opponents predictably 
insist that vaccines harm many and that they don't do all that 
much good. Obviously, such debates could use some good data. 
But, beyond that, we need to recognize that every choice carries 
costs and that we can weigh and choose only among imperfect 
options. Even if we can agree that a vaccine will kill a small 
number of children but will save a great many, how are we to 
incorporate into our decision-making the notion that every 
human life is beyond price? How should we weigh the value of 
a few priceless lives that might be lost if vaccinations proceed 
against the value of many priceless lives that might be lost if vac- 
cinations are curtailed? (Chapter 3 extends this discussion of 

In short, some numbers are missing from discussions of social 
issues because certain phenomena are hard to quantify, and any 
effort to assign numeric values to them is subject to debate. But 
refusing to somehow incorporate these factors into our calcula- 
tions creates its own hazards. The best solution is to acknowl- 

edge the difficulties we encounter in measuring these phenome- 
na, debate openly, and weigh the options as best we can. 


A third category of missing numbers involves what is deliber- 
ately uncounted, records that go unkept. Consider the U.S. 
Bureau of the Census's tabulations of religious affiliation: there 
are none. In fact, the census asks no questions about religion. 
Arguments about the constitutionally mandated separation of 
church and state, as well as a general sense that religion is a 
touchy subject, have led the Census Bureau to omit any ques- 
tions about religion when it surveys the citizenry (in contrast to 
most European countries, where such questions are asked)." 

Thus, anyone trying to estimate the level of religious activity 
in the United States must rely on less accurate numbers, such as 
church membership rolls or individuals' reports of their atten- 
dance at worship services. The membership rolls of different 
denominations vary in what they count: Are infants counted 
once baptized, or does one become an enrolled member only in 
childhood or even adulthood? Are individuals culled from the 
rolls if they stop attending or actively participating in religious 
activities? Such variation makes it difficult to compare the sizes 
of different faiths (as discussed further in chapter 6). Surveys 
other than the census sometimes ask people how often they at- 
tend religious services, but we have good reason to suspect that 
respondents overreport attendance (possibly to make a good im- 
pression on the interviewers). 12 The result is that, for the United 
States, at least, it is difficult to accurately measure the popula- 
tion's religious preferences or level of involvement. The policy 

of not asking questions about religion through the census means 
that such information simply does not exist. 

The way choices are phrased also creates uncounted cate- 
gories. Since 1790, each census has asked about race or ethnici- 
ty, but the wording of the questions — and the array of possible 
answers — has changed. The 2000 census, for example, was the 
first to offer respondents the chance to identify themselves as 
multiracial. Proponents of this change had argued that many 
Americans have family trees that include ancestors of different 
races and that it was unreasonable to force people to place them- 
selves within a single racial category. 

But some advocates had another reason for promoting this 
change. When forced to choose only one category, people who 
knew that their family backgrounds included people of 
different ethnicities had to oversimplify; most probably picked 
the option that fit the largest share of their ancestors. For exam- 
ple, an individual whose grandparents included three whites 
and one Native American was likely to choose "white." In a so- 
ciety in which a group's political influence depends partly on its 
size, such choices could depress the numbers of people of 
American Indian ancestry (or any other relatively small, heavi- 
ly intermarried group) identified by the census. Native 
American activists favored letting people list themselves as 
being of more than one race because they believed that this 
would help identify a larger Native American population and 
presumably increase that group's political clout. In contrast, 
African American activists tended to be less enthusiastic about 
allowing people to identify themselves as multiracial. Based in 
part on the legacy of segregation, which sometimes held that 
having a single black ancestor was sufficient to warrant being 

considered nonwhite, people with mixed black and white an- 
cestry (who account for a majority of those usually classified as 
African Americans) had tended to list themselves as "black.” If 
large numbers of these individuals began listing more than one 
racial group, black people might risk losing political influence. 

As is so often the case, attitudes toward altering the census 
categories depended on whether one expected to win or lose by 
the change. The reclassification had the expected effect, even 
though only 2.4 percent of respondents to the 2000 census opted 
to describe themselves as multiracial. The new classification 
boosted the numbers of people classified as Native Americans: 
although only 2.5 million respondents listed themselves under 
the traditional one-ethnicity category, adding those who identi- 
fied themselves as part-Indian raised the total to 4.1 million — a 
1 10 percent increase since 1990. However, relatively small num- 
bers of people (fewer than eight hundred thousand) listed their 
race as both white and black, compared to almost 34 million 
identified as black. 13 

Sometimes only certain cases go uncounted. Critics argue 
that the official unemployment rate, which counts only those 
without full-time work who have actively looked for a job dur- 
ing the previous four weeks, is too low. They insist that a more 
accurate count would include those who want to work but have 
given up looking as well as those who want full-time work but 
have had to settle for part-time jobs — two groups that, taken 
together, actually outnumber the officially unemployed.” Of 
course, every definition draws such distinctions between what 
does — and doesn't — count. 

The lesson is simple. Statistics depend on collecting informa- 
tion. If questions go unasked, or if they are asked in ways that 

limit responses, or if measures count some cases but exclude 
others, information goes ungathered, and missing numbers re- 
sult. Nevertheless, choices regarding which data to collect and 
how to go about collecting the information are inevitable. If we 
want to describe America's racial composition in a way that can 
be understood, we need to distill incredible diversity into a few 
categories. The cost of classifying anything into a particular set 
of categories is that some information is inevitably lost: distinc- 
tions seem sharper; what may have been arbitrary cut-offs are 
treated as meaningful; and, in particular, we tend to lose sight 
of the choices and uncertainties that went into creating our 

In some cases, critics argue that a failure to gather informa- 
tion is intentional, a method of avoiding the release of damag- 
ing information. For example, it has proven very difficult to col- 
lect information about the circumstances under which police 
shoot civilians. We might imagine that police shootings can be 
divided into two categories: those that are justified by the cir- 
cumstances, and those that are not. In fact, many police depart- 
ments conduct reviews of shootings to designate them as 
justifiable or not. Yet efforts to collect national data on these 
findings have foundered. Not all departments share their 
records (which, critics say, implies that they have something to 
hide); and the proportion of shootings labeled "justified" varies 
wildly from department to department (suggesting either that 
police behave very differently in different departments or that 
the process of reviewing shootings varies a great deal).” 

There are a variety of ways to ensure that things remain un- 
counted. The simplest is to not collect the information (for in- 
stance, don't ask census respondents any questions about reli- 

gion). But, even when the data exist, it is possible to avoid com- 
piling information (by simply not doing the calculations neces- 
sary to produce certain statistics), to refuse to publish the infor- 
mation, or even to block access to it, 16 More subtly, both data col- 
lection and analysis can be time-consuming and expensive; in a 
society where researchers depend on others for funding, deci- 
sions not to fund certain research can have the effect of relegat- 
ing those topics to the ranks of the uncounted. 

This works both ways. Inevitably, we also hear arguments 
that peopleshould stop gathering some sorts of numbers. For ex- 
ample, a popular guide to colleges for prospective students 
offers a ranking of "party schools." A Matter of Degree — a pro- 
gram sponsored by the American Medical Association to fight 
alcohol abuse on college campuses — claims that this ranking 
makes light of and perhaps contributes to campus drinking 
problems and has called for the guidebook to stop publishing 
the list.” While it is probably uncommon for critics to worry 
that statistics might be a harmful moral influence, all sorts of 
data, some will contend, might be better left uncollected — and 
therefore missing. 


Another form of missing numbers is easy to overlook — these are 
figures, once public and even familiar, that we no longer re- 
member or don't bother to consider. Consider the number of 
deaths from measles. In 1900, the death rate from measles was 
13.3 per 100,000 in the population; measles ranked among the 
top ten diseases causing death in the United States. Over the 
course of a century, however, measles lost its power to kill; first 

more effective treatments and then vaccination eliminated 
measles as a major medical threat. Nor was this an exceptional 
case. At the beginning of the twentieth century, many of the 
leading causes of death were infectious diseases; influenzal 
pneumonia, tuberculosis, diphtheria, and typhoidltyphoid fever 
also ranked in the top ten, 18 Most of those formerly devastating 
diseases have been brought under something approaching com- 
plete control in the United States through the advent of vaccina- 
tions and antibiotics. The array of medical threats has changed. 

Forgotten numbers have the potential to help us put things 
in perspective, if only we can bring ourselves to remember 
them. When we lose sight of the past, we have more trouble as- 
sessing our current situation. However, people who are trying 
to draw attention to social problems are often reluctant to make 
comparisons with the past. After all, such comparisons may re- 
veal considerable progress. During the twentieth century, for 
example, Americans' life expectancies increased dramatically. 
In 1900, a newborn male could expect to live forty- six years; a 
century later, male life expectancy had risen to seventy-three. 
The increase for females was even greater — from age forty - 
eight to eighty. During the same period, the proportion of 
Americans completing high school rose from about 6 percent to 
about 85 percent. Many advocates seem to fear that talking 
about long-term progress invites complacency about contempo- 
rary society, and they prefer to focus on short-run trends — es- 
pecially if the numbers seem more compelling because they 
show things getting worse. 19 

Similarly, comparing our society to others can help us get a 
better sense of the size and shape of our problems. Again, in dis- 
cussions of social issues, such comparisons tend to be made se- 

lectively, in ways that emphasize the magnitude of our contem- 
porary problems. Where data suggest that the United States 
lags behind other nations, comparative statistics are common- 
place, but we might suspect that those trying to promote social 
action will be less likely to present evidence showing America to 
advantage. (Of course, those resisting change may favor just 
such numbers.) Comparisons across time and space are recalled 
when they help advocates make their points, but otherwise they 
tend to be ignored, if not forgotten. 


One final category deserves mention. It does not involve poten- 
tially relevant numbers that are missing, but rather includes ir- 
relevant or erroneous figures that somehow find their way into 
discussions of social issues. Recently, for example, it became 
fairly common for journalists to compare various risks against a 
peculiar standard: the number of people killed worldwide each 
year by falling coconuts (the annual coconut-death figure usu- 
ally cited was 150). Do 150 people actually die in this way? It 
might seem possible — coconuts are hard and heavy, and they 
fall a great distance, so being bonked on the head presumably 
might be fatal. But who keeps track of coconut fatalities? The 
answer: no one. Althoughit turns out that the medical literature 
includes a few reports of injuries — not deaths — inflicted by 
falling coconuts, the figure of 150 deaths is the journalistic 
equivalent of a contemporary legend."’ It gets passed along as a 
"true fact," repeated as something that "everybody knows." 

Other legendary statistics are attributed to presumably au- 
thoritative sources. A claim that a World Health Organization 

study had determined that blondness was caused by a recessive 
gene and that blonds would be extinct within two hundred 
years was carried by a number of prominent news outlets, 
which presumably ran the story on the basis of one another's 
coverage, without bothering to check with the World Health 
Organization (which denied the story)."' 

Legendary numbers can become surprisingly well estab- 
lished. Take the claim that fifty-six is the average age at which 
a woman becomes widowed. In spite of its obvious improbabil- 
ity (after all, the average male lives into his seventies, married 
men live longer than those who are unmarried, and husbands 
are only a few years older on average than their wives), this sta- 
tistic has circulated for more than twenty years. It appeared in a 
television commercial for financial services, in materials distrib- 
uted to women's studies students, and in countless newspaper 
and magazine articles; its origins are long lost. Perhaps it has 
endured because no official agency collects data on age at wid- 
owhood, making it difficult to challenge such a frequently re- 
peated figure. Nevertheless, demographers — using complicat- 
ed equations that incorporate age-specific death rates, the per- 
centage of married people in various age cohorts, and age 
differences between husbands and wives — have concluded that 
the average age at which women become widows has, to no 
one's surprise, been rising steadily, from sixty-five in 1970 to 
about sixty-nine in 1988."" 

Even figures that actually originate in scientists' statements 
can take on legendary qualities. In part, this reflects the diffi- 
culties of translating complex scientific ideas into what are in- 
tended to be easy-to-understand statements. For example, the 
widely repeated claim that individuals need to drink eight 

glasses of water each day had its origin in an analysis that did in 
fact recommend that level of water intake. But the analysis also 
noted that most of this water would ordinarily come from food 
(bread, for example, is 35 percent water, and meats and vegeta- 
bles contain even higher proportions of water). However, the 
notion that food contained most of the water needed for good 
health was soon forgotten, in favor of urging people to consume 
the entire amount through drinking. 23 Similarly, the oft-repeated 
statements that humans and chimpanzees have DNA that is 
98 percent similar — or, variously, 98.4, 99, or 99.44 percent sim- 
ilar — may seem precise, but they ignore the complex assump- 
tions involved in making such calculations and imply that this 
measure is more meaningful than it actually is."' 

Widely circulated numbers are not necessarily valid or even 
meaningful. In the modern world, with ready access to the 
Internet and all manner of electronic databases, even figures 
that have been thoroughly debunked can remain in circulation; 
they are easy to retrieve and disseminate but almost impossible 
to eradicate. The problem is not one of missing numbers — in 
such cases, the numbers are all too present. What is absent is the 
sort of evidence needed to give the statistics any credibility. 

The attraction of legendary numbers is that they seem to give 
weight or authority to a claim. It is far less convincing to argue, 
"That's not such an important cause of death! Why, Til bet 
more people are killed each year by falling coconuts!" than to 
flatly compare 150 coconut deaths to whatever is at issue. 
Numbers are presumed to be factual; numbers imply that some- 
one has actually counted something. Of course, if that is true, it 
should be possible to document the claim — which cannot be 
done for legendary numbers. 

A related phenomenon is that some numbers, if not them- 
selves fanciful, come to be considered more meaningful than 
they are. (Chapter 5 also addresses this theme.) We see this par- 
ticularly in the efforts of bureaucrats to measure the unmeasur- 
able. A school district, for example, might want to reward good 
teaching. But what makes a good teacher? Most of us can look 
back on our teachers and identify some as better than others. 
But what made them better? Maybe they helped us when we 
were having trouble, encouraged us, or set high standards. 
My reasons for singling out some of my teachers as especially 
good might be very different from the reasons you would cite. 
Teachers can be excellent in many ways, and there's probably 
no reliable method of translating degree of excellence into a 
number. How can we measure good teaching or artistic genius? 
Even baseball fans — those compulsive recordkeepers and lovers 
of statistics — can argue about the relative merits of different 
athletes, and baseball has remarkably complete records of play- 
ers' performances. 

But that sort of soft appeal to the immeasurability of per- 
formance is unlikely to appease politicians or an angry public 
demanding better schools. So educational bureaucrats — school 
districts and state education departments — insist on measuring 
"performance." In recent years, the favored measure has been 
students' scores on standardized tests. This is not completely 
unreasonable — one could argue that, overall, better teaching 
should lead to students learning more and, in turn, to higher 
test scores. But test scores are affected by many things besides 
teachers' performance, including students' home lives. And our 
own memories of our "best teachers" probably don't depend on 
how they shaped our performances on standardized tests. 

However imperfect test scores might be as an indicator of the 
quality of teaching, they do offer a nice quantitative measure — 
this student got so many right, the students in this class scored 
this well, and so on. No wonder bureaucrats gravitate toward 
such measures — they are precise (and it is relatively inexpensive 
to get the information), even if it isn't clear just what they mean. 
The same thing happens in many settings. Universities want 
their professors to do high-quality research and be good teach- 
ers, but everyone recognizes that these qualities are hard to 
measure. Thus, there is a tremendous temptation to focus on 
things that are easy to count: How many books or articles has a 
faculty member published? (Some departments even selectively 
weigh articles in different journals, depending on some measure 
of each journal's influence.) Are a professor's teaching evalua- 
tion scores better than average? 

The problem with such bureaucratic measures is that we lose 
sight of their limitations. We begin by telling ourselves that we 
need some way of measuring teaching quality and that this 
method — whatever its flaws — is better than nothing. Even if 
some resist adopting the measure at first, over time inertia sets 
in, and people come to accept its use. Before long, the measure 
is taken for granted, and its flaws tend to be forgotten. The crit- 
icism of being an imperfect measure can be leveled at many of 
the numbers discussed in the chapters that follow. If pressed, a 
statistic's defenders will often acknowledge that the criticism is 
valid, that the measure is flawed. But, they ask, what choice do 
we have? How else can we measure — quickly, cheaply, and 
more or less objectively — good teaching (or whatever else con- 
cerns us)? Isn't an imperfect statistic better than none at all? 
They have a point. But we should never blind ourselves to a sta- 

tistic's shortcomings; once we forget a number's limitations, we 
give it far more power and influence than it deserves. We need 
to remember that a clear and direct measure would be prefer- 
able and that our imperfect measure is — once again — a type of 
missing number. 


When people use statistics, they assume — or, at least, they 
want their listeners to assume — that the numbers are mean- 
ingful. This means, at a minimum, that someone has actually 
counted something and that they have done the counting in a 
way that makes sense. Statistical information is one of the best 
ways we have of making sense of the world's complexities, of 
identifying patterns amid the confusion. But bad statistics give 
us bad information. 

This chapter argues that some statistics are bad not so much 
because the information they contain is bad but because of what 
is missing — what has not been counted. Numbers can be miss- 
ing in several senses: a powerful example can make us forget to 
look for statistics; things can go uncounted because they are 
considered difficult or impossible to count or because we decide 
not to count them. In other cases, we count, but something gets 
lost in the process: things once counted are forgotten, or we 
brandish numbers that lack substance. 

In all of these cases, something is missing. Understanding 
that helps us recognize what counts as a good statistic. Good sta- 
tistics are not only products of people counting; the quality of 
statistics also depends on people's willingness and ability to 
count thoughtfully and on their decisions about what, exactly, 

ought to be counted so that the resulting numbers will be both 
accurate and meaningful. 

This process is never perfect. Every number has its limita- 
tions; every number is a product of choices that inevitably in- 
volve compromise. Statistics are intended to help us summarize, 
to get an overview of part of the world's complexity. But some 
information is always sacrificed in the process of choosing what 
will be counted and how. Something is, in short, always miss- 
ing. In evaluating statistics, we should not forget what has been 
lost, if only because this helps us understand what we still have. 


A recent newspaper column by a prominent political 
commentator began: "It is a truism in politics that 
around 40 percent of Republicans will always vote for 
a Republican presidential candidate and about the 
same percentage of Democrats will vote for their party's candi- 
date. The battle is for the middle 20 percent.'" Percentages -I; 
pundit- o. Numbers that appear to be simple can confuse even 
people who are paid to provide insight to the rest of us. And 
there is no shortage of confusing numbers. 

For instance, claims made in the debate over the proposed 
2003 federal tax cut seemed contradictory. The bill's proponents 
declared that the average family's tax reduction would be more 
than #1,000, but the bill's opponents noted that more than half 
of all families would have their taxes cut by less than ffioo. 2 In 

other words, the average benefit would be either a lot (accord- 
ing to those favoring the bill) or a little (according to the oppo- 
sition). Confused? 

Most of us assume that we understand what average means. 
Although many critics bemoan our innumeracy — our discom- 
fort with numbers — Americans actually consume a steady diet 
of familiar statistics that involve averages, percentages, and the 
like , 3 Crime rates, stock market indexes, and batting averages 
are the stuff not only of daily news reports but of routine, every- 
day conversations. The assumption is that we grasp these num- 
bers — and we probably do, more or less. 

However, familiarity can breed confusion. Even apparently 
simple, straightforward numbers can pose traps for the unwary. 
Inappropriate statistics may be offered, or appropriate numbers 
may be used in inappropriate ways. The result is confusion. 
Perhaps we know we're confused (we realize that we don't un- 
derstand the figures), or perhaps we don't (we imagine that we 
understand numbers when we actually do not). Sometimes the 
people who give us a bad number may themselves be confused; 
in other cases, they know what they're doing, and they're trying 
to hoodwink us. 

There are many routes to statistical confusion, and this chap- 
ter cannot hope to discuss more than a few. While understand- 
ing some of the most frequently encountered problems will re- 
quire coming to grips with a few basic mathematical and logi- 
cal principles, our real concern will be exploring how social 
processes — that is, people counting — contribute to these errors. 
The chapter begins with common problems that involve famil- 
iar statistics, such as averages and percentages, and then ad- 
dresses special issues raised by confusing graphs. 


The simplest statistic is, of course, a count — someone tallies up 
a total and reports it: our town has so many residents, its police 
force recorded this many crime reports last year, and so on. 
Counts can be flawed, particularly when the items being count- 
ed are partially hidden (which makes it difficult to get a complete 
count) or when they are very common (which can make count- 
ing so expensive that we must settle for cheaper but less accurate 
estimates). There is also the issue of what counts — it is impor- 
tant to understand how what is being counted is defined and 
measured.’ But, overall, a count seems remarkably straightfor- 
ward. The concept is easy to understand; we've all counted 
things. A count is a single number that corresponds clearly to a 
familiar notion: how many are there? It is difficult to get con- 
fused about a count. Alas, the same cannot be said for other sim- 
ple statistics; even basic arithmetic can inspire confusion. 


One of the most common sorts of arithmetic confusion involves 
the concept of an average. The standard method of calculating 
an average, learned in some half-forgotten arithmetic class and 
usually taken for granted, is to total up scores and divide by the 
number of cases. If a group of children take a 10- word spelling 
test, we can add the number of words each child spelled cor- 
rectly and divide that total by the number of children to give us 
the group's average score — say, 8.2 words spelled correctly. 

The average calculated using this familiar method is techni- 

cally termed the mean. The mean is a useful measure as long as 
the scores do not vary wildly. (A child's score on our spelling test 
cannot be lower than o or higher than io, for instance.) But 
imagine a factory with ninety workers, each earning 840,000; 
nine managers, each earning 880,000; and a chief executive 
officer, who brings home — I am somehow hesitant to write 
"earns" — say, $6 million. We calculate the mean income for the 
people working in our factory as follows: 

90 x 8 40,000 = $3,600,000 income total for workers 

9x8 80,000 = $ 720,000 income total for managers 

1x8 6,000,000 = $ 6,000,000 income total for the CEO 

8 10,320,000 (total income)/ 1 00 (total people) 
= 8103,200 mean income 

This mean is pretty much meaningless. No one at the factory 
earns the average; the nine managers' salaries (880,000) are clos- 
est to the mean, but that average figure (8103,200) is far re- 
moved from either the workers' earnings or the CEO's income. 

One solution to this problem is to present a different measure 
of the average — the median, instead of the mean. The median is 
the middle case in a distribution. To calculate the median, we 
list the cases in the order of their scores, from the lowest to the 
highest, and then take the value of the middle score. In our fac- 
tory example, with one hundred people, the fiftieth and fifty- 
first lowest incomes are in the middle. Both of these incomes are 
840,000, so the median income for our factory is 840,000. In this 
case, the median score gives a more accurate sense of a typical 
income than the mean — after all, 90 percent of the people in 

our factory earn $40,000, making it the typical salary. Because 
income distributions often include figures that vary wildly, the 
median is the preferred measure, used to give a better sense of 
what is average. Thus, we regularly encounter references to the 
"median household income" and so on. 

Whether we choose the mean or the median to express what 
is average, we lose some information. In our factory example, 99 
percent of the people earned less than the mean income, so 
$103,200 is an average only in a very peculiar sense. But using 
the $40,000 median forces us to lose sight of those people who 
make more — in the CEO's case, vastly more. The median is 
probably the preferable figure in this case, but neither measure 
is perfect; no single number clearly conveys exactly how our 
imaginary factory distributes income. And which figure you 
prefer may depend on the point you want to make: using the 
mean might help to emphasize the substantial income generat- 
ed by the factory, while using the median serves to highlight the 
workers' modest incomes. It is not difficult to find examples of 
such differences: in those contradictory tax-cut claims at the be- 
ginning of this chapter, notice how the bill's proponents referred 
to the mean tax reduction as average, whereas opponents point- 
ed to the median figure. Whenever we confront an average, we 
should be able to tell whether it has been calculated as a mean 
or a median. Ideally, we might also consider whether knowing 
the other figure might change our impression of what's average. 


The percentage is probably our handiest statistic: simple to cal- 
culate, incredibly useful, yet almost intuitively easy to under- 

stand. In its simplest form — when dividing a whole into parts — 
it presents few problems. If we are told that about io percent of 
people write with their left hand, then we can calculate that 
roughly go percent use their right hand to write, that there are 
about nine right-handed writers for every lefty, and so on. 

However, because the percentage is such a familiar and use- 
ful tool, it often is used to present somewhat more complicated 
sorts of information. And things don't have to get much more 
complicated before we can become confused. Imagine that we 
do a study of i,ooo adolescents, classifying them as either delin- 
quent or law-abiding, and as either right- or left-handed. Sup- 
pose that we find 810 law-abiding righties, go delinquent right- 
ies, 80 law-abiding lefties, and 20 delinquent lefties. (Before 
someone gets offended, let me emphasize that these are imagi- 
nary data, meant only to illustrate a point.) To help sort through 
those numbers and make sense of them, let's arrange our data as 
shown in Table 1, with cells for each of the four possible combi- 
nations of handedness and delinquency. 

Table 1. Raw Numbers ( fairly confusing) 

Right-Handed Left-Handed 

Law-abiding 810 80 

Delinquent 90 20 

That didn't help all that much, did it? Let's see what happens 
in Table 2, in which we calculate percentages across the table so 
that each row totals 100 percent. (Note that here and in the fol- 
lowing tables I've included the number of people in parentheses 
in each cell, along with the percentages, which allows you to 

check my calculations.) Table 2 clearly conveys the idea that 
right-handers account for substantial majorities of both law- 
abiding and delinquent adolescents. But that's not terribly in- 
teresting, since we already know that right-handers outnumber 

Table 2. Calculating Percentages Across (still confusing) 

Right-Handed Left-Handed Total 

Law-abiding 91% (810) 9% (80) 100% (890) 

Delinquent 82% (90) 18% (20) 100% (110) 

But watch what happens when we calculate the percentages 
down, as shown in Table 3 , so that each column totals 100 per- 
cent. Suddenly, the pattern becomes clear: in our study, the per- 
centage of lefties who are delinquent (20 percent) is twice that 
of righties (10 percent). These percentages help us understand 
the pattern in the numbers. 

Table 3. Calculating Percentages Down (much clearer) 




90% (810) 

80% (80) 


10% (90) 

20 % ( 20 ) 


100% (900) 

100 % ( 100 ) 

Table 2 illustrates what we mean when we say that someone 
has calculated the percentages "in the wrong direction" or "in 
the wrong way." In general, percentages should be calculated so 

that each value of the independent variable (the cause) totals ioo 
percent. In our example, handedness is obviously the indepen- 
dent variable — no one imagines that being law-abiding or 
delinquent can cause you to become right- or left-handed, but it 
is at least conceivable that handedness might somehow affect 
delinquency! Thus, we need to calculate our percentages so that 
the columns for right- and left-handed adolescents each add up 
to ioo percent (as they do in Table 3). 

Deciding which way to calculate percentages requires a little 
thought. Surprisingly often, you can spot people who ought to 
know better presenting percentages that have been calculated 
the wrong way. Sometimes, wrong-way percentages can seem 
impressive. Suppose that someone announces, for instance, that 
a large percentage of alcoholics — say, 60 percent — experienced 
abuse as children. (Once more, I am simply inventing numbers 
for purposes of illustration.) Both the people who make this 
claim and the people who hear it might consider it to be strong 
evidence that childhood abuse affects the chances that people 
will become alcoholics. But alas, this statistic actually gives us 
the percentages calculated the wrong way. 

Let's think this through. One's alcoholism as an adult cannot 
possibly cause one to have been abused during childhood; 
therefore, abuse — not alcoholism — must be treated as the 
cause, the independent variable. What we want to compare is 
the percentage of people who were abused as children and went 
on to become alcoholics with the percentage of people who 
were not abused as children and became alcoholics. That is, we 
should calculate the percentages for those abused and those not 
abused so that each totals 100 percent. If the first figure is 

greater than the second — as it is in Table 4 — the data indeed 
suggest that a history of abuse might affect one's chances of be- 
coming alcoholic. 

Table 4 . Imaginary Data Showing That Childhood Abuse 
Makes Adult Alcoholism More Likely 

Abused Not Abused 

Alcoholic 20% (120) 10% (80) 

Not alcoholic 80% (480) 90% (720) 

Total 100% (600) 100% (800) 

When we hear that most alcoholics experienced abuse as 
children, we tend to assume that people who were abused are 
more likely to become alcoholics, which is what Table 4 shows. 
In this table, the percentage of alcoholics among those abused as 
children (20 percent) is twice as great as the percentage of alco- 
holics among people who were not abused (io percent). (Note, 
too, that my imaginary numbers include 200 alcoholics and that 
60 percent of them — 120 cases — were abused. Thus, these data 
do support our original statement that 60 percent of alcoholics 
experienced abuse.) 

But now consider Tables 5 and 6, which use different sets of 
numbers. In Table 5 , we see that equal percentages of those 
abused as children and those who were not abused become al- 
coholics (20 percent in both instances). These data suggest that 
childhood abuse has no effect on adult alcoholism. And Table 6 
actually shows that people who were not abused as children are 
more likely to become alcoholics (30 percent) than those who 
were abused (20 percent). 

Table 5. Imaginary Data Showing That Childhood Abuse 
Has No Effect on Adult Alcoholism 


Not Abused 

Not alcoholic 

20% (120) 
80% (480) 
100% (600) 

20% (80) 
80% (320) 
100% (400) 

Table 6. Imaginary Data Showing That Childhood Abui-e 
Ma^es Adult Alcoholism Less Likely 


Not Abused 

Not alcoholic 

20% (120) 
80% (480) 
100% (600) 

30% (80) 
70% (189) 
100% (269) 

Despite these notable differences, a close inspection of the top 
rows of Tables 4-6 reveals that, in each case, our original wrong- 
way percentage (60 percent of alcoholics — 120 of 200 — were 
abused as children) remains the same. Although presenting per- 
centages calculated in the wrong direction encourages us to 
imagine that the overall pattern might be the one depicted in 
Table 4, it is important to understand that all three tables, de- 
spite contradictory data, are consistent with that wrong-way 
percentage. This is precisely what's wrong with calculating per- 
centages the wrong way. Such percentages confuse rather than 

Confusion can also result when we use percentages to de- 
scribe a sequence of changes. Suppose we learn that a stock 
index fell 50 percent between 1980 and 1990 but then rose 95 

percent between 1990 and 2000 . Was the value in 2000 greater 
than in 1980? At first glance, it might seem so ("down 50 per- 
cent, but then back up 95 percent, and 95 is way more than 
50 .. .”), but the answer is no. 

Assume that the index's 1980 value was 1,000. A 50 percent 
decline by 1990 would cause the value to fall to 500 (50 percent 
of 1,000 is 500, which we subtract from the original 1,000). The 
95 percent rise between 1990 and 2000 , however, must be meas- 
ured against the 1990 value (95 percent of 500 is 475 , which we 
add to the 1990 value of 500, for a total of 975 — which is less 
than the 1980 figure of 1,000). That is, a percentage change is al- 
ways calculated against the figure at the beginning of the 
change. In describing any series of changes (such as the shifts 
from 1980 to 1990 to 2000), the outcome of one change creates a 
new basis from which the next change is calculated. Thus while 
each single change may seem easy to understand, we need to 
think carefully when we start comparing a series of percentage 

Quick, try this calculation: if our total stock index rose 50 
percent between 1980 and 1990 and then fell 95 percent between 
1990 and 2000 , which value is greater — the one for 1980 or the 
one for 2000? The answer, of course, is again 1980 — and now 
by a huge margin. Again assume that the 1980 value was 1,000. 
A 50 percent increase would cause a rise to 1,500 in 1990, but a 
95 percent fall from 1,500 would produce a figure of only 75 for 
the year 2000 . The lesson is simple: a series of changes expressed 
in percentages creates numbers that aren't really comparable. 6 
People who present information in this way are probably either 
themselves confused or trying to pull a fast one. 

Although the ideas of averages and percentages are familiar 

and seemingly straightforward, they retain the potential to con- 
fuse us. So it should be no surprise that confusion arises even 
more easily as statistical ideas become more complex. 

The Meaning cf Correlation 

One of the most important forms of reasoning occurs when we 
recognize that two things are related ("when I flip this switch 
up, the light goes on; and when I flip it down, the light goes 
off"). This recognition invites us to suspect that one thing may 
cause the other, which lets us better understand our world and 
plan our actions based on what we think we know ("it's too 
dark, so I'll try flipping the switch"). 

Patterned relationships between two things can take many 
forms: every time A goes up, B goes up; every time A goes up, 
B goes down; when A goes up, B is slightly more likely to go up; 
and so on. Philosophers and scientists classify these relation- 
ships as forms of correlation . When we say that A and B are cor- 
related, we are noting some sort of observable relationship be- 
tween them, whether it is a perfect one-to-one correspondence 
(the light goes on every time the switch is flipped up, and only 
when the switch goes up) or only a slight tendency (people who 
were abused as children are somewhat more likely to become 

Such relationships may be causal. We understand, for exam- 
ple, that flipping the switch causes the light to go on. But that 
understanding is grounded on more than observation; we also 
have a theory to explain the relationship (flipping the switch 
closes a circuit that allows electrical current to flow through the 
lightbulb, heating the filament until it gives off light). Our the- 

ory could be wrong, but in this case we know that lots of people 
have tested the theory of electricity, and it predicts so well that 
we have great confidence in it. 

But — and this is the key point — while causality cannot exist 
without correlation, correlation is not itself sufficient to prove 
causality. Just because two things seem related does not mean 
that one causes the other. To return to our imaginary research 
about childhood abuse and alcoholism, suppose that we do find 
that people who were abused as children are more likely to be- 
come alcoholics than people who were not abused (in other 
words, imagine that we have evidence of the sort shown in 
Table 4). Such findings do not constitute proof that childhood 
abuse causes adult alcoholism. It is possible that the abuse- 
alcoholism relationship is spurzous — that is, some third variable 
might cause the variation in both abuse and alcoholism. For 
example, perhaps it is the case that poor families have higher 
rates of child abuse and that people raised in poverty are more 
likely to become alcoholics. If we expand the data in Table 4 by 
also asking whether the people in our study were raised in 
poor families, we might get the (imaginary) results shown in 
Table 7. 

Table 7. Imaginary Data Showing That Childhood Poverty , 
Not Childhood Abuse , Makes Adult Alcoholism More Likely 

Raised Poor 

Not Raised Poor 


Not Abuied 


Not Abuied 


25% (112) 

25% (50) 

5% (8) 

5% (30) 

Not alcoholic 

75% (330) 

75% (150) 

95% (152) 

95% (570) 


100% (442) 

100% (200) 

100% (160) 

100% (600) 

Suddenly, what looked like a strong relationship between 
abuse and alcoholism disappears. In Table 7, we see that child- 
hood abuse has no effect on adult alcoholism, once we take 
childhood poverty into account. This method of analysis is 
called controlling for a third variable, which in this case is child- 
hood poverty. The relationship between abuse and alcoholism 
now seems spurious because the correlation between the two 
variables is in fact explained by a third variable (poverty). 
Again, my point in presenting these imaginary numbers is not 
to endorse some argument about the actual relationship be- 
tween poverty, childhood abuse, and alcoholism? Rather, I 
simply want to demonstrate that, even when we calculate our 
percentages in the right direction, an apparent relationship be- 
tween two variables can vanish into spuriousness. 

In contrast, a genuinely causal relationship is not spurious. 
But this raises a huge logical problem: we can never prove ab- 
solutely that a relationship is not spurious, because it is always 
possible that some unexamined variable, if only we considered 
it, would expose the relationship as spurious. Thus, one can al- 
ways protest that causality cannot be absolutely proven. This 
argument was the defense adopted by the tobacco industry 
when, for decades, it insisted that research showing a relation- 
ship between smoking and lung cancer did not prove that 
smoking caused cancer. Strictly speaking, they were right. Of 
course, by the same logic, we cannot know absolutely that 
flipping the light switch causes the light to shine. 

Any claim that we have identified a causal relationship must 
be examined critically. I have already suggested several ways of 
testing such claims. We can demand some sort of theory, that is, an 
argument about the causal process that connects the two vari- 

ables (electrical current flows through a closed circuit; tobacco 
smoke irritates lung tissues). We can identify likely third variables 
and check to see whether they are sources of spuriousness. In ad- 
dition, we cancompile evidence by doing more studies. Although 
any piece of research has limitations, if we compile many stud- 
ies — each with somewhat different limitations — whose results 
support one another, we begin to have greater confidence in our 
findings. This is why the evidence for the link between smok- 
ing and lung cancer now seems overwhelming. A vast research 
literature exists, based on many different methods — everything 
from tracing the smoking histories of people with lung cancer 
to comparing the proportion of lung cancer deaths among 
smokers and nonsmokers to inducing lung cancer in laboratory 
animals by exposing them to smoke, and so on. While each 
method has its own limitations, studies using all of these meth- 
ods produce results that support the smoking- lung cancer link, 
offering strong evidence for a causal relationship. 

A related problem occurs when people attribute causality 
after the fact. In such cases, we spot a relationship, and because 
one variable logically precedes the other, we assume that it must 
be the cause. This train of thought combines elements of correct 
and fallacious reasoning. The correct reasoning is that a cause 
must occur before its effect; if we can prove that B occurs after 
A, then we know that B cannot cause A. The fallacy is that A is 
not necessarily the cause simply because A precedes B — remem- 
ber that the relationship between A and B may be spurious. 

Most heroin addicts report having smoked marijuana at 
some point before they began using heroin. But can we assume 
that marijuana use in some way causes heroin use? On the one 
hand, the fact that marijuana smoking usually precedes trying 

heroin is not proof of causality — after all, just about every ad- 
dict also ate ice cream at some point before using heroin, and we 
don't peg ice cream as a cause of addiction. But that critique is 
obviously imperfect: marijuana, unlike ice cream, is illegal; and 
we might reasonably suspect that dabbling in one illegal drug 
might foster more serious drug use. On the other hand, the re- 
lationship between marijuana smoking and heroin use is very 
weak: only a small fraction of those who try marijuana become 
heroin addicts. 

Some contemporary critics of drug use try to gloss over the 
issue of causality by declaring that marijuana is a "gateway" 
drug — that is, marijuana use may not cause heroin addiction, 
but it might be a gateway through which most heroin users pass 
on their way to addiction? This analogy, however, is ambigu- 
ous; it does not specify the nature of the link between marijua- 
na and heroin. After all, what is a gateway? Should we envision 
a gate that we could somehow keep closed? In other words, if 
we could keep people from trying marijuana, could we ensure 
that they would not try heroin? Or is the gateway just a well- 
trodden path among a set of alternative routes, so that closing 
the gate wouldn't have much effect? And what should we think 
about those marijuana smokers who do not become addicts? To 
be sure, most heroin addicts have passed through the marijuana 
gateway, but relatively few of the people who go through that 
gateway go on to become addicts. The gateway notion is too 
vague to be much help in understanding drug problems or 
weighing policy options. Thinking about causality needs to be 
less sloppy. 

These issues barely begin to consider the complexities of cor- 
relations between variables. The formal study of statistics — the 

content of most chapters in most statistics textbooks — is devot- 
ed largely to this topic, to measures of the strength of relation- 
ships between variables. (In those textbooks, the term correlation 
also has a narrower, technical meaning, as a particular way of 
thinking about and measuring such relationships.) Such sophis- 
ticated calculations have become far more common, thanks to 
the widespread availability of powerful computers and easily 
mastered statistical software packages. Statistical procedures 
that, thirty years ago, required using one of the large mainframe 
computers available only at a few universities now can be com- 
pleted on a typical student’s standard desktop computer. Today, 
virtually anyone can produce — if not necessarily understand — 
highly sophisticated statistics. This ability has created a contin- 
ual escalation in the complexity of statistical analyses, in an 
effort to specify increasingly complicated relationships among 
ever more variables, by using measures that ever fewer people 
can hope to understand. 

Nonetheless, the basic principles regarding correlations be- 
tween variables are not that difficult to understand. We must 
look for patterns that reveal potential relationships and for evi- 
dence that variables are actually related. But when we do spot 
those relationships, we should not jump to conclusions about 
causality. Instead, we need to weigh the strength of the rela- 
tionship and the plausibility of our theory, and we must always 
try to discount the possibility of spuriousness. 


As we've seen, words — and numbers — can indeed be confus- 
ing. Perhaps it's time to turn to an approach that seems more 

Black hair 

25 % 


50 % 

Brown hair 

FIGURE 1 . A simple pie chart (left) and bar graph (right) represent- 
ing die hair color of an imaginary group of children. 

basic and more easily understood. We sometimes use the word 
envision to refer to our ability to comprehend information. So 
let’s see whether we can’t transform statistics into clearer, visual 

Graphs such as simple pie charts, bar graphs, and line graphs 
are among the most familiar methods of conveying statistical 
patterns. The basic idea is to represent numbers as pictures. The 
essential standard for judging graphs is remarkably simple: an 
accurate display should present visual proportions equivalent 
to the numeric proportions being represented. To take a very 
straightforward example, suppose that we have a group of 
twelve children in which three have blond hair, six have brown 
hair, and three have black hair. We could convey this informa- 
tion in a pie chart or in a bar graph, as shown in Figure i. 

In both graphs, numbers are translated into spatial equiva- 
lents. Half of the children have brown hair, so their slice of the 
pie chart is equal to half; similarly, because twice as many chil- 

4 3 


dren have brown hair as have either blond or black hair, the bar 
for brown hair is twice as tall as the other two bars. In both rep- 
resentations, visual proportions reflect numeric proportions. 

The Damaging Effects of Aesthetics 

Graphs seem so obvious and intuitive that you might think it 
would be difficult to louse them up. In fact, it is surprisingly 
easy, and it has become easier in recent years. In large part, bad 
graphs are driven by aesthetics. People want their graphs to 
seem striking, attention-getting. The graphs in Figure 2, for ex- 
ample, are boring because they don't display dramatic differ- 
ences. In contrast, those in Figure 3, which present exactly the 
same information, seem interesting. 

The two sets of graphs differ only in their vertical scales. 
Both graphs in the first pair (Figure 2; are drawn to show the 
full range of possible values, beginning with zero at the bottom 
and ranging up to a value a bit above the highest number being 
graphed. These graphs have the virtue of keeping visual pro- 
portions true to numericproportions. The problem, of course, is 
that they are hard to read — it is difficult to see much difference 
among the bar graph's bars or much fluctuation in the line on 
the line graph. 

Of course, we almost always want to use graphs to display 
differences or change; it is usually the differences — not the sim- 
ilarities — that tell the story. While in theory the least deceptive 
graph is one with a full vertical scale (as in Figure 2; , that scale 
can obscure the differences. One popular solution is to truncate 
the graph, that is, to cut off the bottom portion of the vertical 
scale, producing graphs such as those in Figure 3. This ap- 

2000 2001 2002 2003 2004 2000 2001 2002 2003 2004 

FIGURE 2 . A bar graph (left) and a line graph (right) with zero as 
die base minimize differences. 

2000 2001 2002 2003 2004 2000 2001 2002 2003 2004 

FIGURE 3 . A bar graph (left) and a line graph (right) whose bases 
have been selected to emphasize differences. 

proach is not necessarily illegitimate, as long as you carefully 
label the values so that readers can understand that the graph 
has been truncated. Truncating the scale has the effect of mak- 
ing the graph more visually interesting; by emphasizing differ- 
ences and changes, it highlights the data’s dramatic qualities. 


Some authorities argue that it is never acceptable to truncate 
bar graphs (because the relative heights of the bars usually con- 
vey the key information), but that it may make sense to trun- 
cate line graphs (with clear labeling) where the focus is the pat- 
tern of changes. Many newspapers, for example, publish daily 
line graphs that display the previous day's stock market fluctu- 
ations; these graphs are truncated and change their vertical 
scales from day to day, to emphasize shifts during the previous 
day's activity. While their different scales mean that the graph 
published on one day cannot be compared to that published 
on the next, the papers' readers presumably understand these 

Several relatively recent changes have magnified the dis- 
ruptive influence of aesthetics in the creation of graphs. One of 
the most important is the widespread availability of graphics- 
producing software. Today, pie charts are everywhere, although 
the old-fashioned method of drawing them by hand with a 
compass, ruler, and protractor is becoming a lost art. Instead, 
creating pie charts with a computer has become so simple that 
people don't give it a thought — and it shows. 

Consider Figure 4. This pie chart accompanied a newspaper 
story about children who had been abducted by family mem- 
bers; it shows the various durations of the abductions. 

The chart is confusing for two reasons. First, the slices of the 
pie are not arranged in any clear order. If we move clockwise 
around the pie, we read: "One week to less than a month," "One 
month to less than six months," "One to six hours," “Twenty- 
four hours to less than a week," and "Other." The sequence 
makes no sense; the chart confuses more than it clarifies. 

Second, the chart contains that peculiar "Other" category. 

Duration of abduction 

- 23 % 

16 % 

hours to 

fess than a One to six 
week hours 

m ^k 


n One monl 

FIGURE 4 . A confusing pie chart depicting the duration of child 
abductions by family members, (Source; Ryan Cormier, A Miscing 
Child, Unbearable Pain,” The News Journal; Wilmington, DE, July 
13, 2003, p, Abused by permission,) 

What does it include? This chart was based on a table published 
in a government report. 9 Examining the original table reveals 
that the 22 percent listed as “Other” in the chart includes the 

Abductions lasting less than one hour 3 percent 

Abductions lasting seven to twenty-four hours 4 percent 

Abductions lasting more than six months 6 percent 

Children not returned but located 6 percent 

Cases for which there is no information 3 percent 

Total 22 percent 

Presumably, the chartmaker grouped these categories together 
because they were all relatively small, but the result is incoher- 
ent: the “Other” category includes some very brief episodes, 
some very long ones, and some about which, apparently, noth- 

b 7 

Confusing numbers 


information n 

an u 

ith || 

— 27% ' 

week — I 

Less than 
one week 


FIGURE 5 . An improved pie chart depicting the duration of child 
abductions by family members. 

ing is known. It makes no sense to lump these cases together in 
a single category. 

Different, and more useful, pie charts could be derived from 
the same data. Figure 5, for example, offers a chart that presents 
the categories in logical order (from brief to lengthy), does not 
jumble together very different cases in the same category, and 
conveys a clear pattern — that abductions of children by family 
members tend to be brief. 

Another problem is that most graphics software packages 
offer a variety of “attractive” display options — for example, you 
can tilt a pie chart to view it from an angle instead of head-on, 
which turns the circular pie into an oval. This may be an aes- 
thetic improvement, but it absolutely undercuts the accuracy 
and usefulness of the chart, because wedges formed by the equal 
angles from the center of an oval need not have equal areas. 
Thus, this method of making a pie chart more interesting and 

4 8 


attractive actually violates the central principle of statistical 
graphics — proportionality of space and numbers — and con- 
veys a distorted, inaccurate impression. Other "improvements" 
have similarly damaging effects, such as showing the edge of a 
tilted pie chart to the viewer (so that it seems to be a three- 
dimensional disk), which exaggerates the visual importance of 
those slices that can be viewed edge-on. 

Similarly, most graphics software packages automatically 
truncate the vertical scales on bar and line graphs to generate dra- 
matic displays that highlight the differences in the data. For ex- 
ample, I had no difficulty producing the graphs in Figure 3; I en- 
tered some imaginary data into a popular spreadsheet program, 
asked for a bar graph and a line graph, and each one popped up 
on my computer screen. In sharp contrast, it took a lot more work 
to produce the full-scale graphs shown in Figure 2, even though 
they involved exactly the same numbers, because I had to cir- 
cumvent the software's default option and enter additional com- 
mands. (I am ashamed to admit that I could not figure out how 
to do this; I needed someone familiar with the program to explain 
the sequence of secret commands.) In short, graphics software 
often makes it simpler to draw distorted graphs than to draw 
proportional ones. As is so often the case in life, one major attrac- 
tion of doing things the wrong way is that it is so much easier. 

Computer software that generates statistical graphics is thus 
a mixed blessing, providing ease of use but almost inviting 
abuse. Because the software offers default formats, it doesn't 
take much thought — or care — to produce a graph. As a result, 
we are bombarded with unnecessary graphs. A full-fledged pic- 
ture may be worth a thousand words, but the information con- 
tained in the typical pie chart can usually be conveyed in a sen- 

tence. And the elaborate menus of bells and whistles offered by 
many software packages, such as tilting a pie chart to expose its 
edge, or turning a graph's bars into three-dimensional figures 
and converting the graph's bottom into a slope, intended to add 
drama, often distort the visual proportions so that the whole 
purpose of the graph — to help people visualize the relative pro- 
portions in the data being presented — is undermined. 

A related phenomenon has been journalists' adoption of pic- 
torial elements to "liven up" graphs. Thus, instead of circles di- 
vided into wedges or simple bars of different heights, graphs 
drawn in this style present odd-shaped figures. Take a look at the 
example shown in Figure 6. The national newspaper USA Today 
helped to popularize this style of graphics, but it is widely used. 
The problem is that, however pleasing to the eye such illustra- 
tions may be, they do a terrible job of conveying information. It 
is often hard to figure out which spatial elements in the figure 
correspond to the numbers they are supposed to represent. One 
of the leading theorists of graphs, Edward J. Tufte, has coined 
the memorable term chart] un\ to refer to all of the extraneous el- 
ements that convey no information and yet litter many contem- 
porary charts and graphs." In extreme cases, chartjunk can make 
it next to impossible to decipher the meaning of a graph. 

Figure 6 is fairly typical of the little feature graphics that ap- 
pear in the lower corners of newspaper pages. This one reports 
the results of an online poll of "self-selected respondents" who 
were apparently asked, "How much will you or do you owe in 
student loans?" Unfortunately, the results of any online survey 
are almost certainly meaningless because the sample is not rep- 
resentative. Not everyone has access to the Internet, only a tiny 
fraction of those who do are likely to stumble across any par- 

By Lori Joseph and Marcy E, Mullins. USA TODAY 

Graduates in debt 

How much will you or do 
you owe in student loans? 

18 % 

15 % — , 

Nothing Less $5,000- $10.000- $20,000- More 
than $10,000 $20,000 $35,000 than 
$5,000 $35,000 

online poll of 5.224 
self-selected I 
respondents May 28 

FIGURE 6 . The meaning of this graphic is obscured by chartjunk, 
{Source: USA Today, August 6 , 2002, p, 1 A; © USA Today, reprinted 
by permission.) 

ticular survey, and only some people (which ones? who knows?) 
will bother completing it. We don’t know whether the sample 
accurately represents the population in question (which, in this 
case, may be all college students or former students, although 
we have no way of confirming that either guess is correct). 
These data, in short, are worthless. 

And that’s the good news. The graph in Figure 6 contains 
bars, represented by stylized greenbacks; the bills seem to be 
sticking out from a graduation cap. (I confess that I puzzled 
over this drawing for quite a while before I figured out what 

5 1 


this was supposed to represent. I was confused by the thing that 
looks like a pencil with an electrical cord, until I realized that it 
was intended to be the tassel for the mortarboard.) But there is 
no way of telling where the bars in the graph begin — some- 
where inside that cap, but where? We can tell that 18 percent is 
greater than 13 percent, which is in turn greater than 11 per- 
cent — but then we already knew that. What we don't get is any 
clear, visual sense of the relative proportions of these quantities, 
because some unknown part of each bar is hidden from us, and 
the uneven, peaked contours of the cap suggest that the ob- 
scured proportion probably differs from bar to bar. 

Making things even more confusing, the viewer's eye is 
drawn to several features on the cap that might — but on in- 
spection prove not to — represent the baseline for the bars. In 
addition to the edge of the cap, we have a curved shadow, the 
cord and tassel, the edge of the mortarboard, and the mortar- 
board's shadow. This is real chartjunk: it makes the viewer 
work to decipher meaning from the drawing's features, efforts 
that will be unrewarded because those features aren't related to 
any information the graph is supposed to convey. In short, this 
graph presents meaningless data in an unreadable form — a 
problem that's increasingly common. Almost any day's newspa- 
per offers examples no better and often much worse, with ir- 
regularly shaped pie charts, bar graphs with bars of indetermi- 
nate dimensions, and so on. 

The combination of aesthetic considerations and computer- 
assisted graphics can make even straightforward, impeccably la- 
beled graphs prepared by professionals unintentionally decep- 
tive. Consider Figure 7 , which reprints bar graphs that first ap- 
peared in a publication of the American Sociological Association. 

figure 3 j 

Percentage of Sexi 

ually Active High School Females by Race 

hitpank f (jjn ak i n on-hi^nic wh I te non- h (spa n k black 

Females females 

Soirtt Abtivert* UtxAlDOOS from 1991 and 1997 Ybutn fWk B^Jvxx Surveyt 

figure 4 

Percentage of Sexually Actrve High School Males by Race 



hrspanic males non-hkpank white norvhispanic black 

Sourct AuthOn' UOuUftom from 1991 and 199? Youth Hnk BcN»«or Suiwn 

FIGURE 7 . A graphic double standard. ( Source: Barbara Risman and 
Pepper Schwartz, "After the Sexual Revolution," Contexts i, no. i 
[February 2002]: 19, © 2002 by American Sociological Association; 
reprinted by permission.) 

In reprinting these bar graphs, I have retained the relative 
proportions found in the original article. The upper graph shows 
how the percentages of high school females who were sexually 
active changed between 1991 and 1997, with the data broken 
down for three ethnic groups (Hispanics, whites, and blacks). 

5 3 


The lower set shows the comparable information for males. 
Each graph, by itself, is clear. But when they are viewed togeth- 
er, the appearance is deceptive. The page layout allots about half 
again as much height for the female graph as for the male graph, 
which consequently has shorter bars. Based on the heights of the 
bars, our eye tells us that males must have been less sexually ac- 
tive than females, even though a close reading of the percentages 
reveals that, in five of the six comparisons, males were actually 
more — sometimes markedly more — sexually active. The effect, 
probably a result of negligence in laying out the page, is to give 
a visual impression exactly the opposite of what the data show. 

And this is a mild example. It is easy to find pictorial displays 
of numeric information that are almost impossible to decipher, 
in which considerations of aesthetics and drama have simply 
swept information aside. Figures 8 and 9, for example, reprint 
two pictorial displays — one hesitates to call them graphs or 
charts — from a recent "atlas of human sexual behavior." 

The graphic shown in Figure 8 uses cloudlike shapes to pre- 
sent data on how often young men and women think about sex, 
with smaller clouds representing smaller percentages. (Why 
clouds? We can't be sure — the original graphic appears over a 
rough map of the southern hemisphere, so perhaps they are sup- 
posed to be clouds in the sky, or perhaps they are meant to evoke 
the cloudlike shapes that cartoonists use to denote unspoken 
thoughts.) But the cloud sizes are not remotely proportional to 
the numbers being represented. For example, 67 percent is 
nearly three and a half times greater than 1 9 percent, but the 67 
percent cloud is many times larger than the 19 percent cloud. 
The graph actually makes it harder to visualize the scale of 
differences among the various numbers. (Although it's difficult 


Number of times 18- to 44-year-olds 
in the USA think about sex 

every day 
or several 

times a day 


a few 
times a week 
or month 

less than 
once a month 
or never 

FIGURE 8 . A graph in which visual proportions are unrelated to the 
numbers being represented, (Source: Judith Mackay, The Penguin Atlas 
of Human Sexual Behavior [New York: Penguin, 2000], p, 21; graphics 
© 2000 Myriad Editions, Ltd,, used by permission of Viking Penguin,) 

to tell, given the irregular shapes of the clouds, the creators of 
this graphic may have committed the classic error of making the 
clouds 3 width and height proportional to die numbers they rep- 
resent. 11 The problem is that our eye does not see width and 
height but total area: if cloud B is twice as wide and twice as tall 
as cloud A, cloud B 3 s area looks four — not two — times larger 
than cloud As.) 

Even worse is the display “ Alcohol Impedes Pregnancy, 33 
shown in Figure 9. The data, which are very simple, show that 
women who drink more have somewhat greater difficulty be- 
coming pregnant. Whereas 64 percent of women who had 
fewer dian five drinks per week became pregnant widiin six 




Percentage of women in Denmark, 
with different weekly alcohol 
consumptions, becoming 
pregnant within six months of 
discontinuing contraception. 

Fewer than five drinks 64% 

More than ten drinks 55% 

FIGURE 9 . A visual display that graphs the scale instead of the data. 
(Source: Judith Mackay, The Penguin Atlas of Human Sexual Behavior 
[New York: Penguin, 2000], p. 47; graphics © 2000 Myriad Editions, 
Ltd., used by permission of Viking Penguin.) 

months after stopping use of contraception, only 55 percent of 
those who reported having more than ten drinks per week be- 
came pregnant within the same period. 

To illustrate these data, we are given rows of wineglasses. 
The 64 percent pregnancy rate among the women who drank 

5 6 


less is represented by five wineglasses, while the 55 percent 
pregnancy rate among heavier drinkers gets ten wineglasses. At 
first glance, this is confusing — why use the smaller image to 
represent the higher pregnancy rate? But then all becomes 
clear: fewer than five drinks gets five glasses; ten or more drinks 
gets ten glasses. This must be a graphic for people (perhaps 
heavy drinkers?) who need help visualizing that five drinks are 
fewer than ten. It harkens back to those cave drawings where 
shepherds who lacked written numbers supposedly kept track 
of their flocks by drawing one sheep for each animal. Of course, 
by choosing to represent the independent variable (that is, the 
number of drinks) rather than the dependent variable (the preg- 
nancy rates), the graphic in Figure 9 abandons any effort to con- 
vey information. 


Overall, aesthetic considerations seem to cause much of the mis- 
chief in contemporary graphs. We should also recognize the 
possibility that graphs' creators may deliberately manipulate 
aesthetics in an effort to slant their presentations, but we need 
not jump to this interpretation. Remember, the standard graph- 
ics software programs adopt default options, such as truncated 
vertical scales, that guarantee distortion. Moreover, many bad 
graphs seem to lack any agenda. Even if we agree that the pie 
chart in Figure 4, the graduation cap in Figure 6, and the clouds 
and wineglasses in Figures 8 and 9 constitute poor graphics 
practices, it is hard to detect a deceptive intent behind those in- 
coherent displays. 

In other cases, however, the choices made about how to pre- 

sent data seem intended to reinforce a particular argument. 
Every graphic — like every statistic — reflects a series of choices: 
What will be shown? How will it be displayed? Some selectiv- 
ity is inevitable, but this necessity can be abused. 

In Figure io, we can see two lines on a graph — a pretty good 
graph — from a government publication on the birth rate among 
teenagers. One line tracks the birth rate (that is, the number of 
births per i,ooo women in the age group) over the second half of 
the twentieth century. Although the line shows some fluctua- 
tions, it is apparent that the birth rate among teenagers ages 
fifteen to nineteen generally declined during this time: it peaked 
in 1957^ at 96.3 births per i,ooo ? but was only 48.7 in 2000, "the 
lowest level ever reported for the Nation?' 12 This might seem 
surprising, given the frequency with which the media carry 
alarmed stories about teen pregnancies and births. 

But consider the second line, which reports the percentage of 
teen births involving unmarried mothers. In 1957, only 13.9 
percent of teen births were to an unmarried teenager, but this 
figure rose to 78.7 percent in 1999- That is, even as the teen birth 
rate has been falling, the percentage of births to unmarried 
teenagers has been rising. The problem is not that the birth rate 
among teenagers has been increasing — it has not. Rather, the 
concern is that a growing share of the births that do occur are to 
unmarried teenagers, who often find it more difficult to support 
and care for their children. In previous decades, couples mar- 
ried earlier (sometimes because the bride was pregnant) ; today, 
marriage tends to be postponed, even when a pregnant woman 
decides to give birth. 

So what should we think about trends in teen births? Neither 
line tells the complete story. We might read the declining trend 




0 1 1 0 

1950 1960 1970 1980 1990 2000 

NOTE: Data for 2000 are preliminary. 

FIGURE 10 . The two lines in this graph tell a complicated story. 
(Source: Stephanie J. Ventura, T J. Mathews, and Brady E. Hamilton, 
"Births to Teenagers in the United States, 1940—2000 ” National Vital 
Statistics Reports 49 , no. 10 [September 25 , 2001]: 2.) 

in the teen birth rate as indicating that things are getting better, 
but at least some critics would view the growing proportion of 
births to unmarried teens as evidence of things getting worse. In 
this case, the numbers are not wrong or deceptive; both lines in 
the graph are based on very good data (federal compilations of 



all reported births). But neither line by itself conveys a clear un- 
derstanding of what's happening; we need to look at the two 
lines together to get a better overall sense of the complex ways 
society is changing. 

But it is easy to imagine advocates who might want to pro- 
mote a particular point of view about teen pregnancy — that it 
either does or doesn't represent a crisis — and who selectively 
choose to present a graph with only one of these lines (the one 
that supports their perspective). Which trend these advocates 
decide to highlight — the declining teen birth rate or the rising 
percentage of births to unmarried teens — is likely to shape how 
we think about this issue. 

Similarly, it makes a difference how much data are dis- 
played. Although Figure io shows that the teen birth rate gen- 
erally declined from 1950 to 2000, we can note that the rate did 
not change very much from roughly 1975 to 1985. The birth 
rate then rose sharply between 1986 and 1991, before falling 
somewhat below the previous lows. Consider how the shape of 
the graph would change if advocates presented data only from 
1975 to 1985, from 1985 to 1991, or even from 1975 to 2000. This 
variation suggests that we should be careful about making too 
much of graphs that display only short-term changes, which 
may turn out to be nothing more than unimportant fluctuations 
in a fairly steady long-term trend. 13 It is always worth asking 
whether data might have been carefully selected to promote a 
particular argument and whether other data exist that might 
support other interpretations. The example of data on teenage 
births ought to remind us that social issues are complex and 
multifaceted, not one-dimensional. 


The sorts of confusion discussed in this chapter are particularly 
unfortunate because they are so unnecessary. Percentages and 
pie charts are relatively simple tools; most of us first encounter 
them in elementary school. Perhaps their very familiarity helps 
lull us into complacency — we assume that we understand com- 
pletely and fail to recognize our confusion. Or perhaps we sud- 
denly realize that there’s something wrong with the numbers 
("that can't be true, can it?! "),but we can't figure out where the 
mistake lies. Confusion fosters frustration, the sense that this 
stuff is just too complicated, which in turn leads to surrender 
("I'll never get it, so there's no point in trying"). 

But we do have an alternative. Instead of declaring ourselves 
powerless, we can spend a few moments trying to understand 
what might be wrong. Is a graph confusing? Examine it. What 
is being represented, and how are those numbers being trans- 
lated into pictures? Do the visual proportions accurately reflect 
the numbers? Is key information missing? What would you 
like to know that isn't shown? Remember that graphs are sup- 
posed to make things clear; if you're confused, it may well be 
the fault of the graph itself. 

Similarly, basic statistics — averages, percentages, and the 
like — should be fairly easy to comprehend. If you're confused or 
shocked by what the numbers show, give some thought to what 
those numbers mean. Where do those figures come from? Who 
produced them, why did they go to the trouble of doing so, and 
how did they go about the task? Would it make a difference if 
the numbers were calculated or presented in different ways? 

It might not be possible to answer some of these questions, 
but even that can be useful information. If we haven't been told 
enough to answer our basic questions, it's a sign that there's 
something wrong. If it seems that the numbers are steering us 
toward a particular point of view, we ought to ask why those 
numbers have been chosen. We can learn to treat confusion as a 
challenge rather than as a sign that we should surrender. 


A series of recent polls asked American adults to estimate 
the percentage of children without health insurance 
and to describe recent trends in the teenage crime rate, 
the teenage birth rate, and the percentage of children 
raised in single-parent families.' A clear pattern emerged: on 
each of these issues, large majorities — between 74 and 93 per- 
cent of the respondents — judged that the problems were worse 
than they actually were. For example, 76 percent responded that 
the percentage of children living in single-parent families had 
increased during the previous five years. In fact, the percentage 
had not changed. Some 66 percent responded that the percent- 
age of teens committing violent crimes had increased during the 
previous ten years, and another 25 percent said that the percent- 
age had remained about the same; but there had actually been a 
decrease. What accounts for this tendency to imagine that things 
are worse than they are? 

Because statistics can be confusing, they make most of us a 
little anxious. In addition, many of the numbers we encounter 
are intended, if not to scare us, at least to make us anxious about 
our world. Of course, most of what counts as newsworthy is bad 
news; our local "happy news" broadcast may end with a forty- 
five-second piece about a skydiving grandmother, but the lead 
story often features a reporter at the scene of a fatal convenience 
store robbery. The same pattern holds for statistics: in general, 
disturbing, scary statistics get more news coverage than num- 
bers reporting good news or progress. It's no wonder we tend to 
exaggerate the scope of social problems. We're used to a fairly 
steady stream of statistics telling us what's wrong, warning that 
things are much worse than we might imagine. 

This tendency to highlight scary numbers reflects the way so- 
cial problems become noticed in our society. Advocates seeking 
to draw attention to a social problem must compete with other 
causes for the notice of the press, politicians, and the public. 
Amid a cacophony of competing claims, advocates must make 
the case that their particular problem merits concern. Their 
claims tend to hit familiar notes: the problem is widespread; it 
has severe consequences; its victims are vulnerable and need pro- 
tection; everyone is a prospective victim; the problem is getting 
worse. Evidence to support these claims often comes from cou- 
pling troubling examples (as discussed in chapter i) with statis- 
tics. Advocates seeking to raise concern naturally find it advanta- 
geous to accentuate the negative; therefore, they prefer scary sta- 
tistics that portray the problem as very common or very serious. 

But advocates aren't the only ones favoring frightening fig- 
ures. The media comb the most routine statistical reports, such 
as the release of census figures, for their most newsworthy — 

usually understood to mean the most troubling — elements. 
And, as we will see in chapter 4, even scientists and officials may 
find that emphasizing scary numbers makes their work seem 
more important. 


When advocates describe a social problem, the statistic we're 
most likely to hear is probably some sort of estimate of the prob- 
lem's size — the number of cases or the number of people affect- 
ed, for example. Large numbers support claims that the prob- 
lem is common and therefore serious. Other statistics, such as 
the number or percentage of people victimized, convey a sense 
of risk; they offer a rough estimate for the likelihood that the 
problem will threaten you or someone you love. These figures 
foster a sense of our vulnerability. Still other statistics, such as 
rates of growth, project the problem into the future, leading us 
to believe that what is now bad is likely to become much worse. 

Such statistics are most compelling when they portray the 
world in especially frightening terms. The more widespread the 
perceived harm and suffering, the more likely it seems that the 
problem will impinge on our world; and the greater the pros- 
pects for things getting worse, the greater our fear. This fear, in 
turn, makes the advocates' claims seem more compelling and 
therefore more likely to influence us. Whereas earlier genera- 
tions of reformers spoke of society's moral obligation to aid its 
most vulnerable and most wretched members, contemporary 
claims often encourage people to act out of self-interest. We de- 
mocratize risk by warning that a problem can touch anyone. 
"NOW NO one is safe from aids" was the message on Life 

magazine's July 1985 cover; another sound bite from the same 
era claimed that "many families are just a couple of paychecks 
away from homelessness." Saying that everyone is vulnerable 
implies that everyone is equally vulnerable; such claims down- 
play well-documented patterns of risk in favor of fostering a 
shared sense of danger. If we see AIDS or homelessness affect- 
ing some other segment of society, then advocates must appeal 
to our sense of moral obligation. But if a threat seems to endan- 
ger everyone, then we all have a vested interest in doing some- 
thing about the problem. It is telling that modern persuasion so 
often invokes self-interest rather than concern for others. 

Even when a problem does not appear to pose a direct, im- 
mediate threat, it is possible to paint a picture of a future when 
things will be much worse. Trends are a way of spotting trou- 
bling patterns; even if things aren't bad now, we may see signs 
that they are deteriorating. Of course, the most frightening 
trends are those that seem to lead inevitably toward catastrophe. 
Statistical estimates for future social problems are hard to con- 
tradict; aside from waiting to see how things turn out, it is 
difficult to debunk a doomsday scenario. Still, a glance at the re- 
cent history of prognostication reveals how cloudy experts' vi- 
sions of the future can be. The popular magazines of my boy- 
hood predicted that the world of 2000 would feature com- 
muters traveling to work in atomic-powered cars and personal 
helicopters, yet they made no mention of personal computers. 
The track record of advocates envisioning the future of social 
problems is not much better: just recall those Y 2 K forecasts of 
the widespread social collapse that would follow the simultane- 
ous failure of the world's computer systems as the calendar 
shifted from 1999 to 2000. 

A tension exists, then, between advocates’ need for com- 
pelling rhetoric — claims that can move others to address some 
social problem — and the limitations of the available evidence. 
Commonly, this is resolved by ignoring those limitations in 
favor of presenting the most powerful message. For activists, 
who believe firmly that their cause is right and who may well 
consider the numbers perfectly reasonable, scary statistics have 
obvious appeal. For the media, scary numbers seem newswor- 
thy, the stuff of good stories. Such numbers thus encounter re- 
markably little resistance. This section examines three sorts of 
figures often used to make social problems seem scary: big esti- 
mates, troubling trends, and apocalyptic scenarios. 

Measuring a Problem's Size 

The simplest sort of scary number estimates the size of a social 
problem — the number of people involved, for example, or the 
cost in dollars. This seems straightforward: we have all counted 
things, so we naturally presume that someone must have count- 
ed something to come up with these numbers. If someone’s 
count has produced a big number, we tend to assume that there 
must be a big problem. 

But social problems are notoriously tricky to count. Cases 
may be hard to identify, and it may be difficult to define and 
measure whatever is being counted. Take recent heavily publi- 
cized claims that preventable medical errors kill between forty- 
four thousand and ninety-eight thousand U.S. hospital patients 
each year. These are remarkably scary numbers, both because 
they seem large and because we go to hospitals in the hope of 
preserving our lives, not ending them. But what, exactly, are 

medical errors that kill — and how might we identify them and 
count them? The fact that we are given a fairly wide range of 
numbers for the death toll reveals that these numbers are esti- 
mates, not precise counts. So how did people arrive at these 

The answer is a little complicated. These particular estimates 
were derived from two studies of hospital discharges that re- 
viewed patients' records to identify "adverse events" (injuries 
caused by medical mistakes); the researchers concluded that 
about 3 to 4 percent of patients experienced such injuries. How- 
ever, neither study measured the percentage of adverse events 
that were preventable or the percentage of preventable adverse 
events that led to death — both of these figures were later esti- 
mated by people who reinterpreted the data from the original 
studies." These later estimates of deaths, not the original re- 
search on adverse events, were the statistics that attracted pub- 
lic attention, despite critics who argued that the basis for those 
estimates was not made clear. 

In addition, the original studies did not consider the overall 
health of each patient. One later study adopted a more refined 
analysis that did consider this factor. The results of this research 
remind us that hospital patients are, after all, often very ill. 
Imagine a patient who is already seriously ill, who is not expect- 
ed to live more than a few days. A medical error — even a "pre- 
ventable adverse event" — might be the immediate cause of that 
patient's death; in fact, the precarious health of such patients 
makes them particularly vulnerable to the effects of medical 
mistakes. But such cases are not likely to be chosen to exemplify 
the danger of medical mistakes. Advocates and the media favor 
more melodramatic examples, pointing to patients who, prior to 

the adverse event, had long life expectancies — for example, a 
high school athlete whose surgery for a minor injury led to se- 
vere brain damage? The study that took into account the overall 
health of each patient suggested that "optimal [that is, mistake- 
free] care.. . would result in roughly i additional patient of 
every io,ooo admissions living 3 months or more in good cogni- 
tive health.'" In other words, these researchers argued, medical 
errors rarely kill patients with good life expectancies. 

The point of this example is not to argue that hospitals don't 
make fatal errors — surely they do. Nor do I mean to dismiss 
some studies and endorse others. The point is that measures of 
a problem's size may not be nearly as straightforward as they 
seem. This example illustrates how tricky it can be to measure 
what might appear, at first glance, to be an unambiguous phe- 
nomenon — patients killed by medical errors. Even assuming 
(optimistically) that we can identify which deaths result from 
medical mistakes, should we count every fatal error? Some 
might answer that, certainly, every patient's death ought to 
count. But others might see a difference between an error that 
shortens the life of a comatose, terminally ill patient by a single 
day and one that robs a relatively healthy young person of 
decades of life. And does our sense of the problem change if we 
discover that cases of the latter sort are relatively rare? 

There are no right answers to such questions; reasonable peo- 
ple can disagree about what ought to count. But such subtleties 
rarely figure into discussions of social problems, given the con- 
siderable rhetorical advantages of depicting a problem as being 
as large — and as scary — as possible. And, of course, using com- 
pelling examples to illustrate the problem can make the figures 
seem even more frightening. 

Troubles with Trends 

Even if a problem isn’t all that large now, it may be growing. 
Measurements over time allow us to identify trends, that is, pat- 
terns of change. This is an important form of reasoning, but, 
again, it is not as straightforward as it might seem. The basic 
problem with assessing trends is maintaining comparable meas- 
urements: if we don't measure the same things in the same way 
on each occasion, our figures may reflect changes in how we 
count rather than changes in anything we are counting? 

One way of testing claims about social trends is to ask what 
might be causing the change. Suppose that the media announce 
that reports of, say, in-law abuse have been rising. Why, we 
should ask, might this be happening? Is there some reason to 
suspect that the number of in-laws involved in abuse is grow- 
ing? Perhaps. But isn't it also possible that people are now pay- 
ing more attention to in-law abuse? (Obviously this is true, as 
there are now news reports about the topic.) Maybe people are 
becoming more familiar with the problem, more likely to deem 
it serious, and therefore more likely to report it; and maybe the 
authorities, in turn, are doing a better job of keeping records of 
those reports. Advocates often dismiss such alternative explana- 
tions; they may argue that giving more attention to in-law abuse 
has created some sort of "backlash," with increased concern 
somehow causing more cases of abuse. Claims beget counter- 
claims, but the burden of proof must fall on those who argue 
that the trend exists. 

A couple of guidelines suggest themselves. First, we should 
be suspicious of claims that trends have suddenly reversed di- 
rection. In general, social patterns change slowly because social 

arrangements have considerable inertia. Social networks are 
webs of connections, reinforced by sets of cultural assumptions. 
Neither those networks nor those assumptions are likely to 
change all at once. When we think about why some people 
commit crimes, it can help to also consider why most people's 
behavior, most of the time, is law-abiding. Criminologists offer 
all sorts of answers, focusing on family dynamics, the state of 
the economy, the nature of the criminal justice system, the mes- 
sages conveyed by the larger culture, and so on. The incidences 
of criminality and law-abiding behavior may well depend on all 
of these. The very complexity of these causal linkages makes it 
harder for trends to suddenly shift: while one cause of crime 
might undergo a dramatic alteration, it is unlikely that all the 
causal factors will change at the same time. Despite this com- 
plexity, however, when people warn about some new trend, 
they tend to argue that a particular change in one specific factor 
is having a dramatic effect. 

Even when new trends do emerge, simple one-variable ex- 
planations probably cannot account for the development. For 
example, after crime rates rose during the 1980s, they reversed 
direction and began falling during the 1990s. Various claims at- 
tributed the new trend to particular causes, such as the war on 
drugs, "broken- windows" policing (that is, strictly enforcing 
laws against public disorder), or more police on the streets. But 
criminologists who sought to investigate and explain the falling 
crime rates concluded that a combination of factors — including 
economic prosperity and changing patterns in drug use — was 
at work. 6 

Second, we should be suspicious of explanations that attribute 
a trend to some sort of anxiety produced by our fast-changing 

society. We do live in a world marked by more or less constant 
change, but this is nothing new. Since the Industrial Revolution 
(usually dated from the first half of the nineteenth century), 
change has been part of Americans' ongoing experience. When 
we marvel at how the Internet has speeded up communication, 
for example ("it's changed everything!"), we forget the dramatic 
transformations wrought by the spread of telephones in the 
twentieth century or the rise of telegraphy in the nineteenth. 
Concerns that the pace of change threatens to disrupt America's 
social fabric have been voiced for at least two centuries. Even 
when we have confidence in our ability to measure trends, we 
need to be wary of jumping to conclusions about their causes. 

Apocalypse Soon ? 

Contemporary discussions of social problems frequently warn 
not only that troubling trends are getting worse but that terrible 
catastrophe awaits. These warnings take many forms: concerns 
about warfare spiraling out of control (nuclear war, nuclear win- 
ter, the hazards of chemical or biological weapons of mass de- 
struction); environmental disasters (overpopulation, resource de- 
pletion, pollution, global warming); medical fears (epidemics of 
new diseases such as HIV or Ebola, medical problems caused by 
pollution); anxieties about economic collapse; and other exotic 
threats, from asteroid collisions and robotics (artificial intelli- 
gences that push people aside) to nanotechnology (engineered 
materials that outcompete biological life-forms) — anddon't for- 
get the Y2K crisis. It is, apparently, a dangerous world out there. 

Needless to say, when apocalyptic visions feature statistics, the 
numbers usually lack precision. Often, the method adopted is 

the one pioneered by Thomas Malthus, the eighteenth-century 
parson who explained that famine was inevitable because popu- 
lation growth must outstrip agricultural production. Malthus's 
model was simple and easily understood; anyone who accepts 
its assumptions must conclude that the outcome — catastrophic 
famine — is unavoidable. The only problem is that Malthus's as- 
sumptions have proven wrong: population growth can be and 
has been controlled in society after society (most experts expect 
global population to stop growing sometime during this centu- 
ry), and agricultural production has in fact expanded faster than 
the population. 

The lesson is that apocalyptic scenarios — and especially those 
that are more than fantastic ("hey, it could happen!") — depend 
on their assumptions. The accuracy of those assumptions has 
everything to do with whether the scenario is worth our worry. 
The world is very complicated, more complicated than the most 
elaborate computer models. Yet, when we talk about social 
problems — even huge problems that might threaten life as we 
know it — we tend to reduce complexity to simplicity. 

I certainly lack the knowledge to assess the scientific basis 
for warnings about global warming — and I suspect that most 
people who work for the news media aren't much better quali- 
fied. We depend on scientific experts to advise us on such mat- 
ters. However, I do know enough — as should the folks in the 
media — to doubt that any single bit of evidence is sufficient to 
establish that catastrophic global warming is occurring. For ex- 
ample, a biologist's report that the range of the Edith's checker- 
spot, a California butterfly, had shifted northward led the press 
to treat this finding as important evidence of the impact of 
global warming. Later analyses questioned this interpretation, 

but the point is that evidence of a change in the habitat of a par- 
ticular butterfly species isn't sufficiently compelling to either 
confirm or discredit the argument that human activity is caus- 
ing potentially catastrophic global warming? Surely there 
ought to be many, many such bits of evidence if claims about 
global warming are true. Yet news media tend to fix on such iso- 
lated reports: the stories are easy to understand (the butterflies 
have moved north); they lend themselves to illustration (we can 
imagine announcers speaking over videotaped butterflies flut- 
tering); and they can be heralded as evidence of a larger, fright- 
ening trend. 

Apocalyptic claims do not have a good track record. And as- 
sertions that statistics support such claims — particularly argu- 
ments that simple, easily understood numbers are proof that 
the future holds complex, civilization- threatening changes — 
deserve the most careful inspection. 


Risk statistics have become one of the most common types of 
scary numbers. We talk about "increased risk," "risk factors," or 
being "at risk." The watershed in our understanding of risk may 
have been the 1960s, a decade that included such landmark 
events as the release of the 1964 surgeon general's report on to- 
bacco and health. While critics had long warned that smoking 
damaged health, the tobacco industry had insisted that no con- 
vincing evidence made this causal link. The surgeon general's 
report had great impact precisely because it seemed authoritative 
(although few Americans could have explained in any detail 
how the surgeon general had drawn the conclusions in the re- 

port) and because it claimed to offer a comprehensive overview 
of a large body of evidence that led to one conclusion: overall, 
smoking increased one's risk of contracting various diseases. 

The surgeon general's report nearly coincided with the publi- 
cation of two other famous risk-centered books. Rachel Carson's 
The Silent Spring (1962) warned that DDT and other chemicals 
threatened the environment, while Ralph Nader's Unsafe at Any 
Speed (1965) attacked the automobile industry's failure to design 
safer cars. These critiques portrayed everyday products — ciga- 
rettes, chemicals, and cars — as posing serious yet largely hidden 
dangers. Such analyses fostered discussions of risk. By the 
decade's end, a consumer rights movement had emerged that 
sought protection against hazardous products, and the environ- 
mental movement had attracted new support by emphasizing 
the dangers posed by pollution. Increasingly, risks were under- 
stood as hidden, perhaps unrecognized, and dangerous — yet po- 
tentially manageable if properly understood, acknowledged, and 
addressed. By warning the public about these risks, the news 
media had a vital role in this process. 

Many of the trappings of modern life — seat belts; auto- 
mobile air bags; bicycle helmets; foods produced without fat, 
caffeine, or pesticides; smoke-free restaurants and workplaces; 
safe sex; daily baby aspirins; assorted medical check-ups — 
reflect our current understanding of, and efforts to minimize, 
various risks. There is a comic quality to some of this, as we try 
to adjust our lives to the latest news story about the latest study. 
Is drinking bad for your health, or is a daily drink beneficial, or 
is it just red wine that's good for you? (Personally, I'm clinging 
to the notion that dark chocolate prolongs life, and if you have 
convincing evidence to the contrary, I don't want to hear it.) 

When we try to translate these words into numbers, we enter 
the realm of probability. A risk is the chance, the probability, 
that something might occur. Thus, when we say that smokers 
have a higher risk of developing lung cancer, we are not saying 
that every smoker will develop lung cancer, nor are we saying 
that no nonsmoker will develop the disease. Rather, the notion 
of increased risk implies comparing probabilities: if X of every 
i,ooo nonsmokers eventually develop lung cancer, and if smok- 
ers develop the disease at a higher rate, then the number of lung 
cancer cases per 1,000 smokers should be markedly higher than 
X. The idea seems simple, but the numbers quickly lead to 

Probability is not well understood. (This explains why casi- 
nos flourish.) We tend to recognize patterns and assume that 
they are meaningful. If we flip a fair coin four times and get four 
straight heads, some people assume that the next flip will be tails 
(because this outcome is somehow "overdue"), while others as- 
sume that it will be heads (because there is a "streak" going). A 
mathematician would say that both assumptions are wrong be- 
cause each coin flip is independent of the others; that is, what 
happens on the next flip is not influenced by what happened on 
the previous flip. After four straight heads, the odds of heads on 
the fifth flip remain fifty-fifty. Should we get a fifth consecutive 
heads, the odds of heads on the sixth flip are still fifty-fifty. If we 
flip a coin a total of six times, we have sixty-four possible se- 
quences of results. Six consecutive heads (HHHHHH)is one of 
those results; HTHTHT is another. We tend to notice the for- 
mer and consider it remarkable, while the latter seems routine, 
but the odds of getting either pattern are exactly the same: one 

in sixty-four. This is not to say that the odds of getting six heads 
are the same as the odds of getting three heads and three tails; 
twenty of the sixty-four possible sequences involve three heads 
and three tails (HHHTTT, HHTHTT, and so on), whereas 
only one of the sixty-four sequences involves six heads. But any 
particular sequence is equally likely to occur, and the fact that 
some sequences seem to form recognizable patterns does not 
make them any more or less likely to occur. 

Once we realize this, we can understand that all sorts of ap- 
parently unusual combinations — the sorts of things we might 
consider remarkable coincidences — can be expected to occur on 
occasion. If about i o percent of people are left-handed, then the 
odds that the next person we see will be left-handed are one in ten 
(or . i ) ? the odds are one in a hundred that the next two people will 
both be left-handed (.1 x .1 = ,oi), and one in a thousand that the 
next three people will be lefties (.1 x .1 x .1 = .001). Despite these 
odds, if we meet lots of people, we will occasionally run into two 
or even three consecutive left-handers. Even rare things can be 
expected to happen — it’s just that they will happen rarely. 

Converting these principles into statistics — risk calcula- 
tions — routinely leads to confusion. Consider the following 
word problem about women receiving mammograms to screen 
for breast cancer (the statements are, by the way, roughly accu- 
rate in regard to women in their forties who have no other 

The probability that one of these women has breast cancer is 
0.8 percent. If a woman has breast cancer, the probability is 90 
percent that she will have a positive mammogram. If a woman 
does not have breast cancer, the probability is 7 percent that she 

will still have a positive mammogram. Imagine a woman who 
has a positive mammogram. What is the probability that she 
actually has breast cancer ? 8 

Confused? Don't be ashamed. When this problem was posed to 
twenty-four physicians, exactly two managed to come up with 
the right answer. Most were wildly off: one-third answered that 
there was a 90 percent probability that a positive mammogram 
denoted actual breast cancer; and another third gave figures of 
50 to 80 percent. The correct answer is about 9 percent. 

Let's look carefully at the problem. Note that breast cancer 
is actually rather rare (0.8 percent); that is, for every 1,000 
women, 8 will have breast cancer. There is a 90 percent proba- 
bility that those women will receive positive mammograms — 
say, 7 of the 8. That leaves 992 women who do not have breast 
cancer. Of this group, 7 percent will also receive positive mam- 
mograms — about 69 cases of what are called false positives. 
Thus, a total of 76 (7 t 69 = 76 ) women will receive positive 
mammograms, yet only 7 of those — about 9 percent — will ac- 
tually have breast cancer. The point is that measuring risk often 
requires a string of calculations. Even trained professionals 
(such as doctors) are not used to calculating risk and find it easy 
to make mistakes. 

Unfortunately, these same doctors may give exactly this sort 
of information about risk to their patients — who have far less 
training, and may be upset in the bargain. A woman who has a 
positive mammogram is likely to be very troubled by that news 
and will probably be even less able to sort through the numbers 
and calculate the overall risk than the physicians were (who, re- 
member, mostly bungled the answer). 

Measuring Ris\s 

But how do we calculate risks? Where do they get those fig- 
ures? This is a tricky question. Ideally, science proceeds through 
experiment. Suppose that we want to learn whether some ac- 
tivity — say, drinking diet cola (something I do often) — poses a 
health risk. We can imagine a fantastic experimental design in 
which we take two randomly assigned groups of children and 
raise them in identical circumstances, except that the experi- 
mental group drinks diet cola and the control group does not. 
We follow them through adulthood into old age and determine 
whether the groups have different sorts of health problems. 
Obviously, it would be impossible to conduct this experiment — 
it would be ridiculously costly in time and money, to say noth- 
ing of its unethical interference with the subjects' lives. For 
these reasons, risk calculations almost never derive from exper- 
iments with human subjects. 

Instead, researchers must devise alternative methods for 
studying risk. For example, they may identify sick people and 
see whether those who are ill report drinking more diet cola 
than people who are well, or they may compare rates of illness 
in communities known to have high and low rates of diet cola 
drinking, or they may conduct experiments in which some lab 
rats drink diet cola and others don't. All of these designs involve 
methodological compromises; they are imperfect ways of deter- 
mining whether diet cola drinkers run greater risks of ill health. 
On the one hand, this is inevitable; every piece of scientific re- 
search contains design limitations. On the other hand, the im- 
perfections in measuring risk are particularly glaring (because it 
is never possible to study humans under strictly controlled, ex- 

perimental conditions), and therefore the results of these analy- 
ses are imprecise and need to be treated with great care. Two 
cautions are particularly important. 

First, research results should not be treated as compelling 
unless they reveal substantial risk. Imagine a study in which 
subjects who drink diet cola are found to be more likely to ex- 
perience a particular disease than subjects who never touch the 
stuff. Since our study cannot possibly have controlled for every 
aspect of these people's lives, we cannot know for sure that 
drinking diet cola caused the difference. To use the term intro- 
duced in chapter 2, the relationship between diet cola and this 
disease may be spurious. For example, we might suspect that 
diet cola drinkers are more likely to be concerned about their 
weight. Perhaps they get less exercise, or eat more, or are genet- 
ically predisposed to weight gain. How can we be sure that their 
health problems are a result of their choice of drink rather than 
a result of one or more of these other factors? We can't be sure. 
Therefore, before we jump to the conclusion that diet cola is the 
cause of the higher incidence of disease among cola drinkers, we 
ought to have fairly strong evidence. 

But what constitutes strong evidence? A common standard 
in this sort of epidemiological research requires that identified 
risks be three times those in the comparison group (that is, 200 
percent greater). (Confused? If X is 5 , then three times X is 15, 
which is 10 greater than 5 . Since 10 is 200 percent of 5 , 15 is 
three times — or 200 percent greater than — 5 .) This is not an ar- 
bitrary standard. Because such research is not truly experimen- 
tal, it is easy to suspect that apparently causal relationships 
might be spurious. And the weaker the relationship, the more 
likely that it is just an accidental finding, particularly if the risk 

being studied is rare. According to statistical theory — too tech- 
nical to explain here — the chances that an apparent relationship 
(involving a rare risk) is not actually valid diminish only when 
the identified risks are at least 200 percent greater. 9 

Understanding even this much gives us a powerful tool for 
evaluating press reports of recent research. Suppose that you 
pick up tomorrow's newspaper and read that a medical journal 
has published a study indicating that diet cola drinkers are 20 
percent more likely to have a specific medical condition. Such a 
sentence will confuse some people, who, for example, may now 
believe that 20 percent of diet cola drinkers will get this disease. 
Actually, this statistic means nothing of the sort. 

Let's assume that, in the general population, 5 people in 10,000 
have the disease. If diet cola drinkers have a 20 percent increased 
risk, there would be 6 cases of the disease among every 10,000 
diet cola drinkers (20 percent of 5 is 1, so a 20 percent increased 
risk would equal 5 t 1, or 6). In other words, what might seem 
to be an impressive statistic — "20 percent greater risk!" — actu- 
ally refers to a very small difference in the real world: 1 addition- 
al case per 10,000 people. (In fact, we can suspect that researchers 
and media coverage favor the wording “20 percent greater" over 
"a 1.2 risk factor," which means the same thing, precisely because 
it makes the result seem bigger and more dramatic.) 

But remember: to be taken seriously, the research ought to re- 
port a 200 percent greater risk. For example, if the rate is 5 cases 
of disease per 10,000 in the general population, the research 
should reveal a disease rate of at least 15 cases per 10,000 among 
diet cola drinkers ( 15 is three times — 200 percent greater than — 
5 ). Is this a reasonable standard? Well, smokers are about 1,900 
percent more likely to develop lung cancer than nonsmokers. 

Any time you read a news story that reports a risk of less than 
three times, or 200 percent, greater, you have every reason to be 
skeptical of the results. 

As a second caution, we should insist on multiple studies. Any 
single study can be mistaken. Scientists know that to test the va- 
lidity of findings, it must be possible to replicate the research — 
to repeat the study and get similar results. (The bubble of excite- 
ment over the reported discovery of cold fusion in 1989 collapsed 
precisely because researchers in other laboratories were unable to 
replicate the reported results.) It also helps to triangulate re- 
search, that is, to study a phenomenon using different methods. 
Although any one method has its own flaws, the different flaws 
in the various methods can cancel out one another. The link be- 
tween smoking and lung cancer, for example, is considered well 
established because it has been consistently supported in studies 
that use a variety of methods. 

Sometimes researchers compare the results of several studies 
in what is called meta-analysis. The logical assumption is that if 
several studies consistently show an effect, even if the effect is not 
powerful (that is, the risk is less than the 200 percent greater 
standard), the multiple consistent results ought to give us more 
confidence that the relationship is real. The problem with this 
logic is that researchers often do not seek to publish — or have 
greater difficulty publishing — disappointing results. This publi- 
cation bias means that it is hard to get studies with weak results 
published. Thus, meta-analyses tend to include only the most 
successful studies — those with results strong enough to get pub- 
lished. While the meta-analysis technique is not illegitimate, nei- 
ther does it provide particularly strong support. A meta-analysis 

of several studies showing, say, 20 percent greater risk should not 
fill us with confidence in the results. 

It also helps to put risks in some larger context. Every time 
we get in a car and drive to work, we take a risk. We all under- 
stand that traffic accidents kill people. To some degree, we can 
minimize our risk by obeying the traffic laws and wearing our 
seat belts, but the risk never becomes zero, although the chance 
of being killed on any particular journey is very low. Still, such 
routine risks — the sorts of things we take for granted — may be 
far greater than the highly publicized risks that suddenly be- 
come the focus of public attention. When we are frightened, we 
tend to focus on what scares us rather than on the actual risk of 
our being affected, a reaction that has been termed "probability 
neglect.’ 510 

We can see a good example of this in the public's alarmed re- 
action to the news that a sniper was killing people in the region 
around Washington, D.C., during the fall of 2002. Because our 
ordinary, day-to-day assumption is that the risk of being shot by 
a sniper is zero, the news that some risk existed frightened peo- 
ple. Still, in a region containing millions of people, the risk of 
being shot remained very low. Even during the weeks when in- 
dividuals died at the hands of the sniper, people were at much 
greater risk of dying in traffic accidents in greater Washing- 
ton — yet traffic deaths were not headline news. Following the 
mundane advice we've heard all our lives — don't smoke, wear 
seat belts, eat sensibly, and exercise — is likely to increase our life 
expectancies far more than ducking to keep out of a sniper's 
sights or avoiding that food additive that figures so prominently 
in this week's headlines. 

The Rts\ of Divorce 

Another reason that the notion of risk leads to confusion is that 
we're not always sure how best to calculate risks. Consider an 
apparently simple question that turns out to be somewhat com- 
plicated: what proportion of marriages end in divorce? No offi- 
cial agency keeps track of particular marriages and is therefore 
able to identify precisely which ones end in divorce — which is 
the sort of information one would like to have to answer this 
question. Lacking complete and perfect data, analysts are forced 
to use the numbers that are available. Since filing a marriage li- 
cense and obtaining a divorce are both legal steps, official agen- 
cies do keep records of these events, and various jurisdictions 
tally the marriages and divorces they record. Therefore, ana- 
lysts have long divided the number of divorces during a partic- 
ular year by the number of marriages during that year to get a 
rough measure of the likelihood of marriage ending in divorce. 
Since roughly 1960 , the number of divorces has been nearly half 
that of marriages, and commentators often refer to this as the 
"divorce rate." 

The problem is that when we speak of a rate, we are usual- 
ly dividing some number of events (such as deaths or crimes) by 
the population at risk. Thus, both death rates and crime rates 
are usually presented as the number per 100,000 people in the 
population; for example, the FBI reported that the murder rate 
was 5.5 murders for every 100,000 people in the United States 
in 2000 . But who makes up the population at risk when we try 
to calculate a divorce rate? Obviously, it does not include only 
those who married during the same year; in fact, we know that 
relatively few couples get divorced during the calendar year in 

which they marry. Rather, the population at risk is all married 
couples — a very large number indeed. If we calculate the rate 
of divorce by dividing the number of divorces during a partic- 
ular year by the total number of married couples, regardless of 
the length of their marriages, then the divorce rate must be far 
less than 50 percent. All manner of commentators have made 
this point, insisting that marriage is therefore a more stable 
institution and divorce is less common than we might have 

But let's examine this assertion. Imagine a community that 
records two marriages each year — and one of those new mar- 
riages ends in divorce during that same year. In this case, it is 
true that half of all new marriages end in divorce; yet it is also 
true that, with each passing year, the total number of married 
couples will grow by one. Thus, after the first year, dividing the 
current year's lone divorce by the total number of married cou- 
ples will produce a rate lower than 50 percent in spite of the fact 
that half of marriages end in divorce. This reasoning suggests 
that the standard critique used to dismiss high divorce rate sta- 
tistics must be flawed. 

Clearly, measuring the risk of divorce is a tricky problem, 
one that requires both careful thought and, it turns out, a lot 
of data. In 1996, investigators interviewed a very large sample, 
nearly seventy thousand people at least fifteen years old, living 
in some thirty-seven thousand households. The respondents 
were asked about all marriages and divorces in their personal 
histories. For instance, one person might report marrying once, 
forty years earlier, and remaining married to the same spouse; 
whereas another respondent, currently unmarried, might re- 
port marriages in 1970 and 1985 that ended in 1980 and 1992, 

respectively.” These data allowed the investigators to identify 
cohorts of marriages that had occurred during different periods 
(for example, first marriages that took place in 1945-1949) and 
to calculate the proportion of marriages in each cohort that had 
ended in divorce by 1996 . (It is always possible that a couple still 
married at the time of the interview could later decide to di- 
vorce.) Although these data are not complete, because they 
come from a sample rather than from the population as a whole, 
the sample is a good one — about as good as samples get — and 
the data give a glimpse of what happens to particular marriages 
over time (which was, remember, the sort of data we wished for 
at the beginning of this discussion). 

Alas, these data suggest that about half of current marriages 
can be expected to end in divorce. The researchers found im- 
portant cohort differences that reveal how society has changed; 
basically, people in each cohort were likely to have remained 
married longer than those in the cohort that followed. Thus, 
only about 34 percent of sixty-year-old men had had their first 
marriage end in divorce, but the comparable figure for fifty- 
year-old men was 40 percent. Of the women who first married 
during 1945-1949, 70 percent were still married thirty years 
later; but among those whose first marriage occurred during 
i960— 1964, only 55 percent (just over half!) remained married. 
It is too soon to tell what proportion of couples first married 
during 1980-1985 will celebrate their thirtieth anniversaries, 
but we can make projections based on the record so far: only 73 
percent of the women who wed during those years were still 
married ten years later, compared to the 90 percent of those first 
married in 1945-1949 whose marriages lasted at least ten years. 
Based on these data, the investigators projected that, while a 

larger proportion of earlier marriages remained intact, about 
half of recent marriages will indeed end in divorce. 

Thus, answering an apparently simple question — what is 
the likelihood that a marriage will end in divorce? — turns out 
to be a fairly complicated matter. But this sort of complexity is 
glossed over in media reports that glibly report on the risk of 
this or that — an observation that should give us pause. It is all 
too easy to be frightened by risk statistics. We need to keep in 
mind the difficulties of calculating risks as we digest today's 
warning about a newly discovered threat. 


Most often, scary numbers warn that our world is changing. It 
can be unsettling, even frightening, to think about change, par- 
ticularly since media reports tend to focus on changes that are 
for the worse. One of the most useful ideas when considering 
the meaning of change is the notion of trade-off — that is, every 
change involves both costs and benefits. It is impossible to make 
a fair comparison between what came before and what follows 
unless we consider the comparable costs and benefits. 

One of the classic methods of promoting a specific change is 
to contrast the costs of what we have now with the benefits the 
proposed change will bring; similarly, change can be resisted by 
emphasizing the benefits of the existing situation and the 
prospective costs that will be imposed. To protect ourselves — to 
make a fair comparison — we need to compare apples and ap- 
ples. In other words, if we consider today's benefits relevant, we 
must compare them to future benefits, and today's costs ought to 
be compared to future costs. 

It is surprisingly easy to forget to do this. Many critics have 
become suspicious of technological change and point to its costs. 
The comparison, which is often implicit, harkens back to an 
idyllic past when people somehow lived in harmony with na- 
ture, when life was simpler and better. This view through the 
mists of time is a little fuzzy; the critics can see the benefits of the 
past but have trouble making out the costs it entailed. Thus, they 
calculate the costs of, say, deaths caused by air pollution from 
modern power plants, but they forget to tally the death toll from 
indoor pollution caused by cooking over woodstoves. The critics' 
comparison usually involves weighing present or future deaths 
caused by change against a past in which, somehow, death is 
taken for granted. The opposite error occurs when boosters 
highlight the benefits of a proposed change and ignore its costs. 

Comparisons that ignore trade-offs, along with big estimates, 
frightening trends, apocalyptic scenarios, and ill-defined risks, 
are among the most common ways of making statistics scary. 
Because scary numbers are compelling, and because we often 
have difficulty sorting out relative risks and trade-offs, a pes- 
simistic presumption that things must be getting worse runs 
through many contemporary discussions of social problems. 

These gloomy warnings contrast with the lived experiences 
of most Americans. I don't want to imply that every individual's 
world gets better every day in every way; our society features 
plenty of hardship and suffering. However, on average, Ameri- 
cans are living longer than their ancestors, they are healthier 
and better educated, and they have higher standards of living.'" 

There is, in short, a gap between our sense that our own lives 
are going pretty well and our perception that the larger society 
is beset by troubles. This gap regularly appears when public 

opinion polls ask pairs of questions about individuals' own ex- 
periences and their perceptions of the state of the nation. People 
tend to be reasonably satisfied with the teaching provided by 
their local schools but deeply concerned about the quality of 
American education; they often think pretty well of their local 
congressional representative but view Congress as a sinkhole; 
and they report being pleased with the directions their own lives 
are taking, even as they worry that society is on the wrong path. 
Presumptive pessimism colors our thinking about the larger so- 
ciety. As a professor, I have read thousands of term papers and 
examination essays over the years, and I realize that many stu- 
dents simply assume that crime (or poverty, or teen suicide) is 
getting worse, regardless of whether the actual crime rates are 
rising or falling. It is as though we all think of ourselves as liv- 
ing comfortably in Lake Wobegon (among all those above- 
average children), even while we are confident that the larger 
society is headed to hell in a handbasket. 

In recent decades, we have been exposed to a variety of apoc- 
alyptic scenarios, warnings that life as we know it could end. 
Some threats have faded (remember the 1970s fears about a new 
ice age?), but we continue to hear about plenty of paths to ex- 
tinction: nuclear winter, global warming, overpopulation, epi- 
demic disease, economic collapse, terrorism. Scary statistics 
have an important place in these claims. Isolated findings, such 
as the report that a species of butterfly has shifted its habitat, can 
be presented as significant harbingers of impending disaster. 
Even good news can be interpreted as foretelling catastrophe: if 
crime rates are falling, the situation can't last; and should they 
stop falling, it is surely a sign that crime is about to swing back 
out of control. Our readiness to speculate about the largest pos- 

sible implications of small developments means that we are con- 
stantly being warned that big things are in the offing. And, once 
more, we find ourselves frustrated by what seem to be contra- 
dictory claims — alcohol harms your health; no, a glass of red 
wine is good for you; no, a little alcohol in any form is good for 
your heart (but bad for your liver). 

Scary numbers flourish because they are an integral part of 
the way we talk about social life. Advocates of different causes 
seek to scare us because, they insist, we face real threats and be- 
cause we need to be jarred out of our comfortable complacency. 
They are less likely to acknowledge another consideration — the 
competition for our attention. We are surrounded by advocates 
for different causes, each group trying to get us to focus on a 
particular problem. Each cause hopes to stand out from the oth- 
ers. Frightening people is not the only way to win this competi- 
tion, but it often works pretty well, especially if advocates can 
point to statistics to justify the fear. We have every reason to ex- 
pect that scary numbers will remain a key feature of how we 
talk about social problems. These numbers aren't going to go 
away; all we can do is try to approach them with skepticism, to 
assess whether fear is really necessary. 


A couple of times each month, I receive an e-mail mes- 
sage from the editor of some scholarly journal, asking 
whether I'd be willing to review a manuscript. Most 
people know that professors are under pressure to 
"publish or perish." Peer review is a largely hidden part of that 
publication process. Typically, after completing their research 
and writing reports about their findings, scholars submit their 
manuscripts to journals that specialize in publishing such articles. 
The editors of these journals receive more — sometimes far, far 
more — manuscripts than they can possibly publish, and they use 
peer review to help them choose among the submissions. The 
editor sends a copy of each manuscript to several reviewers (in 
sociology, the leading journals usually send copies to three or four 
reviewers). As the term suggests, these reviewers are supposed to 
be the researcher's peers — professionals knowledgeable about 
the topic and therefore qualified to judge the quality of a research 

report. Reviewers may disagree, but a journal's editor will almost 
always reject a manuscript that gets mostly negative reviews. 

Authors, particularly those who have recently had manu- 
scripts rejected, sometimes doubt the integrity of the review pro- 
cess. Some, for example, question whether reviews ought to be 
anonymous (they usually are, although authors and reviewers 
can often guess each other's identity); others raise suspicions that 
a negative review may reflect a reviewer's personal or political 
disagreements with an author. But the peer review process en- 
dures because it seems to work better than any other method for 
selecting the best scholarship for publication. It probably works 
best in the most prestigious journals. There are thousands of 
scholarly journals, but scholars in the various disciplines, spe- 
cialties, and subspecialties recognize that some journals have far 
more readers than others; and competition to publish in the most 
widely read, and therefore most prestigious, venues is intense. 
Such journals are presumed to be especially selective. 

Peer reviewers are gatekeepers. Their job is to identify a 
manuscript's flaws and call them to the editor's attention. Does 
an author seem unfamiliar with other recent research on the 
topic? Did the author choose questionable methods to conduct 
the study? Has the author used inappropriate techniques to an- 
alyze the research results? Reviewers' doubts on such points 
warn editors against publishing weak papers. 

When people refer to a journal as "authoritative," they are 
speaking to the integrity of the journal's review process. Editors 
and reviewers cannot possibly oversee the entire research pro- 
cess and vouch for the accuracy of every word in a manuscript, 
but they can weigh what they read, be alert for warning signs, 
and allow only what seem to be the strongest papers to appear 

in print. Of course, mistakes occur. Over time, some published 
results are called into question; rarely, there is a dramatic ex- 
pos&of scientific fraud — charges that researchers deliberately 
fudged their results to get their work published. But, overall, 
the system seems to work pretty well. 

The authority of social institutions depends on such arrange- 
ments to ensure integrity. We design checks and balances, re- 
quire officeholders to swear oaths, encourage ethics of profes- 
sionalism, and devise other techniques to keep institutions and 
the people who fill them in line. To the degree that we have con- 
fidence in these arrangements, we can place our trust in, among 
other things, the statistics these institutions produce. When we 
are young children, most of us learn to be skeptical of claims (in- 
cluding statistics) that appear in advertisements; we come to ex- 
pect them to be one-sided and distorted. In contrast, we have 
more confidence in statistics produced by scientists or govern- 
ment agencies. Such information is considered more authorita- 
tive because these institutions are presumed to be more profes- 
sional, more impartial, and more committed to the accuracy of 
their numbers. 

Thus, we can speak of authoritative numbers, statistics pro- 
duced by those thought to be authorities, that reach us via insti- 
tutional channels that seem to vouch for the accuracy of the 
figures. In general, these statistics avoid the clumsy errors dis- 
cussed in earlier chapters. Numbers produced by authorities 
rarely involve mistakes in calculation; the methods of collecting 
and presenting the data are ordinarily appropriate. By the time 
most such statistics reach the public, they have been examined 
by colleagues, peer reviewers, editors, and others. These num- 
bers are about as good as statistics get. 

Nonetheless, even authoritative numbers need to be handled 
with care. This chapter examines some examples of statistics 
found in professional journals and government reports, in an 
effort to identify some of the sorts of questions that should be 
asked about such numbers. It begins with an extended discus- 
sion of an article in a major medical journal, a product of the 
peer review process. 


Each morning, I read the Wilmington News Journal. It is the 
principal newspaper in Delaware, but Delaware is not a big 
state, and other newspapers are much bigger than the News 
Journal. Still, it is a fairly typical contemporary newspaper. On 
April 25, 2001, a front-page News Journal story summarized an 
article that had appeared in that week's issue of the Journal of the 
American Medzcal Association , noting that "nearly one of every 
three U.S. children in sixth through ioth grades have been bul- 
lied, or bully other students themselves." Nor was this item 
unique. Two months later, the News Journal ran a story about 
another report in the Journal of the American Medzcal Association 
(more familiarly known as JA MA ) headlined: "Sexual Solicita- 
tion Reported by 20% of Kids Who Use Web." And two months 
after that, a News Journal headline reported on yet another 
JAMA article: ” i in 5 Girls Abused by a Date, Study Suggests."' 

The News Journal can't afford to pay a reporter to read 
through each week's issue of JAMA to locate newsworthy sto- 
ries, so how does my local newspaper get these items? The an- 
swer is that it relies on wire services. But that raises another 
question: how do the wire services cover developments in sci- 

ence and medicine? JAMA, for instance, sends out press releas- 
es about articles in the current issue that its editors hope will 
prove newsworthy. (Not all scientific journals do this; JAMA s 
principal rival for top medical journal honors, the New England 
Journal cf Medicine , does not issue press releases to publicize its 
articles, although it does make advance copies of each issue 
available to the media.)" 

Presumably , JAMA wants to get its name before the public, to 
give people the sense that it publishes important research. 
Among other things, JAMA s visibility makes top researchers 
more eager to submit their research to the journal; publication 
there offers an opportunity to bring one's work to the notice of 
not only fellow professionals but also the larger public. And re- 
searchers who successfully place their papers in highly presti- 
gious journals in turn please their funders — the government 
agencies or private foundations that supply the grants that pay 
for large-scale research. Knowing that their grants led to high- 
ly visible publications confirms to the funders that they spent 
their money wisely. 

It is important to appreciate that this is an extremely com- 
petitive process. Funding agencies winnow through many grant 
applications to select those projects worthy of support. Would- 
be authors submit about ten times more manuscripts to JAMA 
than that journal can publish, and its editors must not only 
choose among these submissions but also decide which articles 
merit press releases. Newspapers are flooded with press releas- 
es and must determine which ones will run in the limited avail- 
able space. By the time a piece of research finds its way into even 
a short item in the News Journal, it has survived several stages at 
which rejection is more likely than selection. 

Remember that the News Journal published reports about 
JAMA articles indicating that bullying affected about 30 percent 
of students, that 20 percent of Internet-using youths had been 
sexually solicited, and that 20 percent of high school girls had 
been violently abused by dating partners. As an experiment, 
imagine that each of those articles had portrayed the problem it 
discussed as being one-tenth — or even one-third — as common; 
that is, imagine that bullying affected between 3 and 10 percent 
of students and that Internet sexual solicitations and dating vi- 
olence each affected between 2 and 7 percent. The findings now 
seem less impressive, don't they? Would the News Journal still 
have published articles about those studies? Possibly — even 
probably — not. Would JAMA's editors have circulated press re- 
leases for articles with those findings? Again, probably not. In 
fact, with those less impressive results, we can suspect that 
JAMA’s editors might have been less likely to publish those arti- 
cles, that the authors might have been less likely to submit their 
papers to such a highly selective journal, and that the funding 
sources would have been less impressed with the reception 
given the published results. In other words, we can imagine that 
everyone in the publication process — the editors at the News 
Journal and at JAMA, the researchers, and the funders — might 
well prefer studies that produce more impressive numbers. 

Let me be clear: I am not suggesting that anything fraudu- 
lent is involved in this process. True, rare scandals reveal that 
researchers have faked their results, but that is not what I'm de- 
scribing. Rather, I'm simply suggesting that there are advan- 
tages to presenting research findings in terms that make the re- 
sults seem as impressive as possible. A report depicting a big 
problem will be favored in the competition to gain attention. 

So let's see how big numbers can be produced. Consider that 
study of bullying. The article in JAMA , "Bullying Behaviors 
Among U.S. Youth," presented results from a large representa- 
tive sample of students (nearly sixteen thousand young people) 
in grades six through ten. The authors were associated with the 
National Institute of Child Health and Human Development, 
which supported the survey. The report received considerable 
coverage by print and broadcast news media, which featured 
the finding that nearly 30 percent of youths "reported moderate 
or frequent involvement in bullying." Researchers have con- 
ducted many other studies of bullying, but few have involved 
samples so large and well drawn. Given the composition and 
size of the sample, and the article's appearance in an especially 
prestigious journal, we might take it as representing the best 
work on the subject. 

What, exactly, is bullying? According to one federal publica- 
tion, "Bullying can take three forms: physical (hitting, kicking, 
spitting, pushing, taking personal belongings); verbal (taunting, 
malicious teasing, name calling, making threats); and psycho- 
logical (spreading rumors, manipulating social relationships, or 
engaging in social exclusion, extortion, or intimidation).” 3 Any- 
one with clear memories of junior high school who reads that 
definition might be surprised that only 30 percent of the respon- 
dents in the JAMA study felt affected. Bullying is a term both 
broad and vague, and much of what might be classified as bul- 
lying is probably fairly common behavior. 

Still, the proportion of students in a survey who report being 
involved in bullying will depend on the questions they are 
asked. The section dealing with bullying in the JAMA article's 
questionnaire began with an explanation: 

Here are some questions about bullying. We say a student 
is being bullied when another student, or a group of students, 
say or do nasty and unpleasant things to him or her. 1 1 is also 
bullying when a student is teased repeatedly in a way he or she 
doesn't like. But it is not bullying when two students of about 
the same strength quarrel or fight, (emphasis in original ) 4 

The students were then asked how frequently they bullied 
others or were bullied during the current school term. Separate 
questions covered bullying in and out of school, although those 
responses were combined for the JAMA article, which did not 
specify how much of the reported bullying occurred in schools. 
For each question, possible answers described different fre- 
quencies of involvement: "I haven't...," "once or twice," 
"sometimes," "about once a week," and "several times a week." 
In presenting their results, the authors defined responses of at 
least weekly experiences as frequent involvement in bullying 
and responses of "sometimes" as moderate involvement. 

These categories form the basis of the study's central finding, 
that nearly 30 percent of youths had moderate or frequent in- 
volvement in bullying. The authors conclude that "bullying is a 
serious problem for U.S. youth" and that "the prevalence of bul- 
lying observed in this study suggests the importance of preven- 
tive intervention research targeting bullying behaviors.” 5 In 
other words, bullying is widespread, and something needs to be 
done about it. 

But does the article demonstrate that bullying is a wide- 
spread, serious problem? The key statistic — that 30 percent of 
youths are involved in bullying — depends on three manipula- 
tions, three methodological choices. First, students could be in- 
volved either as a bully (13.0 percent acknowledged that they 

were bullies), a victim (10.6 percent), or both (6.3 percent). 
Choosing to count bullies as well as victims — that is, all of those 
"involved" in bullying — made a big difference; if the authors 
had chosen to count only the victims, their findings would have 
focused on about 17 percent of students (10.6 t 6.3 = 16.9), not 
on 30 percent. 

Second, the authors included both "moderate" bullying (oc- 
curring "sometimes," that is, more than once or twice during 
the term but less than weekly) and "frequent" bullying (occur- 
ring at least weekly). Adopting a narrower definition would 
have made the findings seem less dramatic; only 8.4 percent of 
the respondents reported being targets of frequent bullying, not 
17 percent. 

Third, remember that the authors combined responses for 
questions about bullying in and outside school. Although the re- 
searchers did not report these data, at least some of those re- 
sponding that they were frequently bullied might have identi- 
fied this as happening only away from school. If so, even fewer 
than 8.4 percent would have reported frequent bullying in 
school. In other words, the authors made a series of choices that 
allowed them to estimate that bullying significantly affected 30 
percent of students. Different choices — say, looking only at vic- 
tims of frequent bullying in schools — would have produced a 
figure only about a quarter as large, if that. 

The point is not that this is a bad piece of research, nor is it 
to deny that bullying may sometimes have serious conse- 
quences. (Some reports alleged that the shooters in heavily pub- 
licized school shootings were reacting to being bullied.) But the 
numbers that emerge from social research must be interpreted 
with care. The finding that 30 percent of students are involved 

in bullying needs to be understood not as some sort of absolute 
fact that has its own independent existence but rather as a prod- 
uct of a particular set of methodological decisions. How the sur- 
vey's questions were worded, the order in which questions were 
asked, and the choices made in interpreting and summarizing 
the results for publication all shaped the findings. Similar meth- 
odological choices affected the well-publicized findings in the 
JAMA articles about Internet sexual solicitations and dating 

It is also important to understand the concerns that can un- 
derpin such research. An anti-bullying movement has arisen, 
which believes that bullying is a serious but neglected problem, 
one that must be addressed. Without such an expression of con- 
cern, the federal government might not have funded this costly, 
large-scale research. Of course, no well-established pro-bullying 
lobby exists; no one argues that bullying is desirable. Therefore, 
we can expect that most researchers studying the topic will seek 
to demonstrate that bullying is a serious problem — and that 
journal editors will prefer to publish articles that support that 

There was nothing dishonest or unprofessional about the 
JAMA piece. Anyone who reads the article will find all of the in- 
formation I've presented in this discussion. But any article must 
be condensed to create an abstract or a press release; only a few 
of an article's many findings are highlighted when the piece is 
summarized. Emphasizing the 30 percent figure made this arti- 
cle seem more newsworthy, while other, less dramatic findings 
were ignored or downplayed in the press coverage. For exam- 
ple, the JAMA piece reveals that the percentages of students who 
reported that they had experienced bullying fell drastically as 

the students aged: 13.3 percent of sixth-graders but only 4.8 per- 
cent of tenth- graders said they experienced frequent bullying. 
Thus, bullying declines as youths mature — hardly a surprising 
finding, but one that might have implications for urgent calls 
for anti -bullying measures. 

In exploring how the results of this study found their way 
onto the front page of my local newspaper, I mean to highlight 
the role of choices in shaping how research gets reported. All re- 
search is a product of a long series of choices. Analysts must de- 
termine what they want to study — a decision that may reflect 
such considerations as their own intellectual interests, their 
sense of what their colleagues consider worthwhile research, 
and the availability of funding. They must also make all man- 
ner of methodological choices: how to draw a sample and col- 
lect data, how to define and measure concepts, how to analyze 
and interpret the results. Research choices are constrained by 
what is already known and by the sorts of time, money, person- 
nel, and other resources available for the study. But these choic- 
es are always consequential; they inevitably shape the results. 
Thus, every study has limitations; one can always argue that, 
had the analysts made different decisions, the findings might 
have been different. This is why scientists insist on both repli- 
cating research (repeating a study to confirm the results) and 
compiling bodies of findings from studies based on different 
choices. As the number of studies with consistent results grows, 
confidence in those findings swells. 

But most of us do not closely follow the gradual expansion of 
scientific knowledge. Rather, we get our information about 
scientific advances from summaries of single studies that ap- 
pear, say, on the front page of our daily paper. And the journal- 

ists who bring us those reports make choices, too: given all the 
stories competing for coverage, and given the limited number of 
newspaper column inches (or broadcast minutes) available, 
which stories merit coverage? With such constraints, the steady 
development of scientific knowledge doesn't seem especially 
compelling to reporters and editors, whereas an apparently 
pathbreaking piece of research seems like news. The news 
media look for drama or human interest. An article reporting 
that bullying is very common seems like a good story because a 
large share of the news audience may find the story relevant to 
their lives — audience members have children, or at least know 
children, and this makes the research seem interesting. (Of 
course, journals that issue press releases for their articles need to 
be aware of the media's concerns; a good press release should 
focus journalists' attention on a study's newsworthy aspects.) 

To complicate matters, scientists have their own agendas: 
they generally want their research to appear in print, to receive 
recognition, and to lead to rewards such as tenure, promotion, 
and further grants. Some commentators tend to equate re- 
searchers' agendas with political ideologies, worrying that stud- 
ies are designed to support liberal or conservative positions. But 
this is only a small part of the story. Researchers may also be al- 
lied with particular theoretical or methodological schools with- 
in their disciplines; for example, bitter debates may occur be- 
tween factions favoring competing statistical models for inter- 
preting research results. Such allegiances and concerns — large- 
ly hidden and incomprehensible to outsiders — often shape re- 
searchers' choices. 

The scientific literature is supposed to be self-correcting. Sci- 
entists understand that no study is perfect, but they believe that, 

over time, the research process will produce a body of findings in 
which we can place our confidence. When researchers have 
reservations about a study's results, they may ask to examine the 
data and offer an alternative analysis, or they may decide to con- 
duct a new study. Slowly, agreed-upon — that is, authoritative — 
knowledge emerges within the research community. But this 
process is slower and more complex than the way most of us con- 
sume the fruits of scientific research, via short news briefs that 
relay the contents of press releases. 


Even the most professionally compiled data can be subject to 
misinterpretation. Consider death records. In the United States, 
the law requires completion of a death certificate for each 
known death; in each case, someone with authority, such as a 
physician or coroner, is expected to assign a cause of death. 
These records travel through bureaucratic channels until they 
eventually find their way to the National Center for Health 
Statistics. The NCHS, in turn, issues an annual report, Vital 
Statistics d the United States , summarizing records of births and 
deaths. The Vital Statistics reports once took the form of three 
Phonebook- size volumes filled with huge, multipage tables; 
now the reports are in electronic form, and anyone can access 
them at the NCHS Web site. The federal government has been 
compiling these records for a long time, and most of the bugs 
have been worked out of the system. Counting births and 
deaths is relatively straightforward, and these data are about as 
complete and accurate — as authoritative — as any we might 
hope to find . 6 

And yet, misadventures are still possible. Consider reports of a 
late twentieth- century rise in suicides among African American 
teenagers. In 1998, the U.S. Centers for Disease Control and Pre- 
vention (CDC) reported that the suicide rate for African Ameri- 
cans between the ages of ten and nineteen more than doubled be- 
tween 1980 and 1995 and that this increase was far greater than 
the increase among white youths./ This was a disturbing finding. 
Teen suicide strikes most adults as especially tragic, as an act ex- 
pressing isolation, despair, and desperation. In addition, most 
Americans would like to think that race relations haven't been 
getting worse. Why would the suicide rate among black teens be 
rising near the end of the twentieth century? 

Reporters who picked up on the CDC report tried to explain 
the apparent trend by contacting psychiatrists and clinical 
psychologists, the professionals usually considered the relevant 
experts; in effect, they own the suicide problem in our culture. 
These authorities mentioned a "post- traumatic slavery syn- 
drome [that] can manifest itself in a range of self-destructive be- 
havior." They also cited "family breakdown, low economic op- 
portunities, undiagnosed depression, unacknowledged grief 
from neighborhood violence, even the additional stress of en- 
tering the middle class," in that "upwardly mobile black fami- 
lies may lack traditional family and community support .” 8 The 
CDC report had suggested: "Black youths in upwardly mobile 
families. . . may adopt the coping behaviors of the larger socie- 
ty in which suicide is more commonly used in response to de- 
pression and hopelessness .” 9 In other words, plenty of after-the- 
fact explanations were offered. 

It is also possible, however, that the increase reflects not a 
change in black youths' behavior, but a change in how their 


Ol T T T T T T 

1970 1 975 1980 1985 1 990 1995 1999 

FIGURE 11. Suicides among African Americans ages ten to nine- 
teen, 1970—1999. (Source: National Center for Health Statistics.) 

deaths are processed by the authorities. 10 Consider Figure 11, 
which traces the number of suicides among African Americans 
ages ten to nineteen from 1970 through 1999. (In order to make 
the stages of my argument as clear as possible, I have chosen to 
graph the actual numbers of recorded deaths rather than the 
death rates. Presenting rates would not change the patterns in 
the data relative to the point of my argument.) As the graph in 
Figure n indicates, we are not talking about a lot of cases: 
according to the NCHS (which, remember, tries to record all 
deaths), there were slightly more than 100 suicides by black 
teens in 1970 and about 250 in 1995, whereas the 2000 census 
identified 6.2 million African Americans between the ages of 
ten and nineteen. 

Figure 12 duplicates the first graph, but it adds a second line, 
showing the numbers of deaths in this age group that are listed 
as having an undetermined cause. In the NCHS listing, “unde- 
termined” is a residual category for those few deaths that are 
not assigned a more specific cause. The comparison of the two 




FIGURE 12. Suicides and deaths resulting from undetermined 
causes among African Americans ages ten to nineteen, 1970-1999. 
(Source: National Center for Health Statistics.) 

lines is striking: as the number of suicides rose during the late 
twentieth century, the number of deaths resulting from unde- 
termined causes fell. 

Figure 13 adds another dimension to the analysis. It shows 
recorded deaths over the same period that were attributed to 
four accidental causes: drowning, gunshots, poison, and falls. I 
selected these four categories of fatal accidents because they are 
also common ways of committing suicide; for example, when an 
official completes a death certificate for a youth who drowned, 
the death might be recorded either as an accident or as a suicide. 
Figure 13 reveals that all four forms of accidental death de- 
clined among black youths during the last decades of the twen- 
tieth century. 

Finally, Figure 14 combines all of the information from the 
three previous graphs into one bar graph. Each bar is broken 
into three segments: suicides on the bottom, deaths resulting 
from undetermined causes in the middle, and the four cate- 




FIGURE 13 . Accidental deaths resulting from drowning, shooting, 
poisoning, and falls among African Americans ages ten to nineteen, 
1970—1999. (Source: National Center for Health Statistics.) 

FIGURE 14 . Deaths from four accidental causes, deaths from unde- 
termined causes, and suicides among African Americans ages ten to 
nineteen, 1970-1999. (Source: National Center for Health Statistics.) 

gories of accidental deaths on top. The total height of each bar 
represents, then, the total number of deaths attributed to all 
these causes. Again, we see that the number of suicides among 
black youths rose during this period, even as deaths from all of 



the other causes fell; overall, the total number of deaths from all 
of these causes dropped by more than half. This is important, 
because all of these deaths represent incidents that might con- 
ceivably be classified as suicides (depending, of course, on other 
available information). If anything, it is probably more difficult 
to classify a death as a suicide than to assign some other cause; 
because family members are more likely to resist a finding of 
suicide, it should be much easier to classify a death as accidental 
or even as a result of undetermined causes. 

In other words, the rise in teen suicides among African 
Americans may not be all that mysterious. Its roots may reside 
not so much in, say, the psychological pressures on youths in up- 
wardly mobile African American families as in a shift in the 
way officials handle the deaths of black teenagers. Whereas such 
deaths once may have been treated as relatively unimportant — 
perhaps brushed off as an accident or as a result of undeter- 
mined causes — officials may now conduct more careful investi- 
gations to arrive at the more difficult designation of suicide.” 
Obviously, it is impossible to prove that this is the correct expla- 
nation for the rising numbers of suicides; that would require re- 
viewing the evidence used to assign cause of death in thousands 
of cases, and most of that information is probably long lost by 
now. But this example does remind us that even the best, most 
complete, most authoritative data — such as the NCHS death 
records — cannot speak for themselves. 

Rather, numbers must be interpreted. In this case, someone 
at the CDC noted a rise in the number of deaths classified as sui- 
cides among African American teenagers and assumed that this 
must reflect a real increase in suicidal behavior. Such behavior 
in turn needed to be explained by identifying changes in the 

youths' lives that made them more suicidal. But an increase in 
the number of deaths classified as suicides need not reflect more 
acts of self-destruction; it might also reveal changes in the way 
officials classify deaths as suicides. Because the reported suicides 
were drawn from apparently authoritative official records, most 
commentators failed to question the rise, even though they 
needed convoluted explanations to account for it. Besides, scary 
statistics about race are common enough that many simply pre- 
sume that they are correct. 

Even apparently straightforward recordkeeping can prove to 
be extremely challenging. Consider a second example: the effort 
to compile the death toll from the September 1 1, 2001, terrorist at- 
tacks on New York's World Trade Center. Airline records made 
it possible to count and name the people who had been on the two 
jets almost immediately. But how many people died in the col- 
lapse of the buildings? No one keeps a master list of the people in- 
side a skyscraper at any givenmoment. Even during the course of 
a normal working day, those present — employees in their offices, 
customers, visitors — form a large, constantly shifting population. 
And the airliners crashing into the buildings led many thousands 
of people to evacuate the towers, even as hundreds of firefighters, 
police, and other emergency personnel entered the structures. 
Moreover, when the buildings fell, the destruction was so com- 
plete that many bodies vanished without a trace. 

In this case, counting the dead turned out to be very compli- 
cated. Within a few days, officials had compiled various lists of 
people reported missing. The names came from firms who 
offered lists of employees thought to have been in the buildings 
and from worried friends and family members who hadn't 
heard from people they suspected might have been in the Trade 

Center. Reports continued to arrive until September 24, when 
the list peaked at 6,453 names. 

Then the list began to get shorter. Officials began to cull du- 
plicate names (for example, a dozen different reports had been 
made for the same woman, each giving different addresses or 
contact numbers). People who had been reported missing 
turned out to be alive (more than fifteen hundred foreigners ini- 
tially reported missing by embassies were located). Investiga- 
tions also identified some seventy fraudulent reports from peo- 
ple hoping to collect survivors' benefits. A handful of names 
were added — for example, people who had been moved to out- 
of-state hospitals before they died from injuries caused by the 
attacks. On September n, 2002, the total had fallen to 2 , 801 , 
which still included 35 to 40 people for whom there was no 
definitive evidence that they had — or had not — died. Even 
after a year of painstaking investigation, the total was not yet 
certain; and, in fact, it continued to change.'" 

The death toll became a subject of contention, particularly 
during the fall of 2001. For a few weeks, some officials continued 
to repeat early estimates of 5,000 or 6,000 deaths, and their rhet- 
oric seemed to argue that these heavy losses were the justification 
for retaliation against the terrorists. Some even criticized the ini- 
tial press stories that predicted (correctly) that the final death toll 
would prove to be much lower. Of course, the horror of the at- 
tack was not somehow proportionate to the numbers lost; the 
final death toll proved to be about half what was originally esti- 
mated, but this did not make the tragedy only half as great. 

Even as some officials tried to carefully tally the casualties, 
others disseminated another dubious statistic about the mag- 
nitude of the catastrophe: they claimed that the World Trade 

Center attack had orphaned 10,000, or even 15,000, children, 
many of whom would need adoption. This estimate could not 
pass even the most casual examination. Even if we take the peak 
estimate for the death toll (6,500), 15,000 orphans would have 
meant that each victim averaged more than two children. More- 
over, if we use the conventional meaning of orphan — a minor 
child who has lost both parents — it is obvious that this claim 
was most improbable: many victims' children would have been 
adults; not all victims would have had children; and most mar- 
ried victims would have been survived by a spouse who could 
continue to care for their children. While thousands of family 
members suffered the loss of loved ones, New York's family 
service officials could not identify a single child of those killed 
in the attack who required adoption or foster care, 13 The World 
Trade Center attack was a terrible event, but it was still possible 
to circulate statistics that exaggerated the extent of the damage. 

I have chosen to focus on death statistics because they seem so 
straightforward; it is far easier to count deaths than to measure 
poverty, unemployment, crime, and most of the other things 
officials count. Official statistics are often the most complete, the 
best — the most authoritative — figures we have, but that does 
not mean that they are perfectly accurate. 

Officials have considerable advantages in collecting statistics. 
Compared to the research projects conducted by scientists, 
many official agencies have generous budgets, which allow 
them to pay people to collect, compile, analyze, and interpret 
data. Compliance with such data collection efforts may be re- 
quired by law; citizens are supposed to cooperate with the cen- 
sus, and birth and death records are mandatory. As data go, 
official statistics tend to be relatively complete. 

But official records are products of the political system and 
therefore are inevitably shaped by political considerations. Every 
decision to collect official information can be a focus for political 
debate. What information do we need? Precisely which infor- 
mation should we collect? How should we collect it? How 
should it be compiled? Which results should be made available? 
How should they be made available, and to whom? What sorts 
of resources should we devote to this process? It costs time and 
money to collect information, so we can assume that someone 
considers the collection effort to be worth the cost. In some cases, 
there may be widespread agreement that the information ought 
to be collected, that this serves some general interest; most peo- 
ple probably approve of keeping birth and death records, for ex- 
ample. People may even agree about what should be counted 
and how. 

Very often, however, matters are more complicated, with 
competing interests trying to shape statistics. Chapter i, for ex- 
ample, noted that ethnic minorities tend to advocate collecting 
census data about ethnicity in ways that maximize their groups' 
numbers. Or take the case of the Consumer Price Index. The 
CPI is widely used as a basis for calculating cost-of-living raises 
for union contracts and government benefits. This makes the 
method of calculating the CPI a matter of more than academic 
interest. Employers and government programs that must pay 
employees based on changes in the CPI favor calculations that 
minimize the growth of the index, whereas those whose earn- 
ings or benefits are tied to CPI increases favor calculations that 
maximize CPI growth. Economists who suggest ways of alter- 
ing the CPI formula to reflect changes in the way people live 
(such as adjusting for the impact of home computers or cell 

phones) find their work criticized not only on intellectual 
grounds but also for its political implications.” 

And, of course, officials' views of their role may vary. At one 
extreme, officials may see themselves as impartial professionals, 
collecting statistics in an unbiased manner. At the opposite ex- 
treme, officials may consider themselves active agents for some 
faction, such as the current political administration, and they 
may deliberately try to produce statistics that support its poli- 
cies. (Note that this need not involve fraud or outright decep- 
tion. It can simply take the form of choosing to count particular 
things or of publicizing particular numbers and emphasizing 
their importance.) Most officials probably fall between these ex- 
tremes; they seek to do a competent, accurate job, yet sometimes 
find their work shaped by their own commitments or by politi- 
cal pressures from others. 


Authoritative statistics depend on our confidence in the institu- 
tions that collect them. Accountants, for example, certify that a 
firm's financial records are in good order, which assures in- 
vestors that they have the information necessary to make wise 
investment decisions regarding that firm. The 2001—2002 reve- 
lations that Enron and other major corporations had adopted — 
and their accountants had approved — various dubious financial 
arrangements produced a major scandal that not only ruined the 
firms directly involved but also threatened investor confidence 
in the larger economy. The federal government subsequently 
passed a corporate reform law requiring that the chief executive 
officers of major corporations personally certify, under penalty 

of criminal sanction, that their firms' records were legitimate. In 
other words, because one layer of institutional protections had 
proven insufficient, the solution was to devise yet another layer 
of reassurance, in order to further guarantee the accuracy and 
reliability of financial recordkeeping. Confidence in authority 
depends on such symbols. 

Such guarantees may seem to be fragile social contracts. 
Ordinary people cannot check or replicate the numbers pro- 
duced by scientists, officials, accountants, and other authorities; 
the costs in time and money would be impossibly high. Instead, 
we rely on the professionalism of those authorities, on their 
pledge to meet the expectations of their clients, the law, their 
peers, and themselves to produce the best possible numbers. 
Statistics from poorer countries that lack the resources to sup- 
port data collection and analysis are often little better than 
guesses , 15 but a rich society expects — and largely receives — 
high-quality statistics from its authorities. In the United States, 
bad statistics are scandalous. Recall the shock when, in the af- 
termath of the 2000 election, people began to understand that 
even mechanized systems of counting votes can lead to errors 
(for example, by failing to count ballots with hanging chads); 
similarly, reports of scientific fraud, officials maintaining inac- 
curate records, or serious accounting lapses become major news 

Despite our expectations, the examples in this chapter 
demonstrate that authoritative statistics have their limits. Data 
collection is never perfect; the "dark figure" of hidden, un- 
counted cases is always present . 16 Every analysis involves choos- 
ing what to count and how to go about counting, and those 
choices always shape the resulting numbers. Often, those 

choices reflect pressures on the authorities. In some cases, all the 
pressure may come from one direction, leading everyone to sup- 
port the same set of choices; but in other instances, competing 
pressures come from those who hope, say, for a big number, 
while others would prefer a smaller figure. And, of course, au- 
thorities have expectations for one another: scientists use the 
peer review process to improve the quality of published re- 
search; accountants have generally agreed-upon standards for 
evaluating accounts; and so on. Inevitably, even the most au- 
thoritative statistics reflect all of these social processes. People 
can and do disagree about the best way to conduct the census 
or measure unemployment or assess the danger of bullying. 
Counting, even when it produces authoritative statistics, is a so- 
cial process. 

In short, the question to ask about any number — even those 
that seem most authoritative — is not "Is it true?" Rather, the 
most important question is "How was it produced?" If some 
numbers are more authoritative, it is because we have more 
confidence in the processes that brought them into being. But 
this is not to say that we should imagine that any numbers offer 
magical solutions to our problems. 


A nyone who follows the news hears about economic 
recessions. In good times, commentators speculate 
about the risk of a recession beginning; in bad times, 
they wonder whether the current recession is about to 
end. It turns out that the authority to make these determina- 
tions, to identify when recessions begin and end, belongs to the 
Business Cycle Dating Committee of the federal government's 
National Bureau of Economic Research. This usually anony- 
mous committee made news in the summer of 2003, when it 
proposed changing the criteria used to determine when a reces- 
sion was ending.' 

The committee had been using several monthly indicators of 
economic activity, including payroll employment (the number of 
people employed in payroll jobs), to identify when recessions 
began and ended. The formula had not included a measure of 
gross domestic product (GDP, the value of goods and services 

produced in the United States) because GDP was measured 
quarterly, not monthly. Because previous recessions had been 
marked by declines in both jobs and GDP, failing to include GDP 
made little difference in designating a recession's start and finish. 

The recession that began in late 2001, however, broke this 
pattern: thanks to improved productivity, GDP began to rise in 
late 2002, yet payroll employment continued to decline. Because 
the committee's formula relied on the jobs measure, which was 
still falling, the official assessment was that the recession was not 
over, even though many observers believed that the economy 
had bottomed out months earlier just before GDP began to rise. 
Therefore, the committee decided to incorporate monthly esti- 
mates of GDP into its calculations, a step that led to a declara- 
tion that the recession had ended, although the committee ac- 
knowledged that there were continuing losses in employment. 

Once again, we see the impact of people choosing what to 
count. Under the committee's old formula, the 2003 economy was 
still in a recession; under the proposed new formula, the recession 
would be over. The committee translates numbers — in this case, 
economic measures — into official labels for the state of the econ- 
omy. In doing so, the committee gives those numbers importance. 

Our culture depends on numbers, and therefore treats them 
seriously. Even when we suspect that our statistics are flawed, 
we realize that we can't get along without figures. The econo- 
my — and the rest of our world — is too complicated to compre- 
hend without resorting to numbers; we need statistics to give us 
a basis for understanding what's happening and for making 
choices. Counting and measuring can help us decide what to do. 
When our attention is drawn to some new social problem, one 
of our first impulses is to quantify it, to measure its scope. 

Statistics, we say, will let us "get a handle" on the problem, as 
though translating the problem into numbers will somehow 
give us the means to bring it under control or at least show us 
how we might achieve control. We act as though numbers have 
amazing powers to illuminate, to make the right choices appar- 
ent — as though they ha we magical properties. 

Magical numbers, then, are figures we imagine to be accurate 
and authoritative, numbers that promise to make our problems 
understandable and therefore manageable. Magical numbers 
seem to transform ambiguity into certainty, to provide a basis 
for complicated decisions. They offer a standard against which 
we can assess the world. At least this is what we tell ourselves. 

This suggests that we should watch for magical numbers to 
appear at our culture's fault lines — at those spots where conflict, 
uncertainty, and anxiety seem particularly intense, where we feel 
the need for a firmer foundation on which to base our actions. 
When someone draws attention to a social problem, for exam- 
ple, it forces us to confront claims that our society doesn't work 
as well as it should, that something must be done to make things 
better. Our culture aspires to perfectibility: we will, we insist, 
"leave no child behind"; we declare war on poverty, on drugs, 
even on cancer. Given these lofty aspirations, drawing attention 
to a social problem is a critique that seems to require action. Of 
course, some people may question whether this problem really 
needs attention or may disagree about the appropriate solutions. 
It is no wonder that such debates over social issues almost always 
feature statistics. Advocates often resort to numbers to bolster 
their claims, to make them seem more certain. Remember, our 
culture presumes that statistics are factual; numbers suggest that 
someone has measured the problem and understands its dimen- 

sions. Figures can make us feel less confused about what we 
ought to do. 

Numbers may be unnecessary in unambiguous situations. 
When people's actions are governed by ritual, by the orders of 
those in command, or by shared moral standards, there is less 
room for choice, for uncertainty or anxiety. But our modern 
world is characterized by complexity and diversity, by compet- 
ing claims and shifting standards. Uncertainty is common, and 
we often turn to statistics for their magical ability to clarify, to 
turn uncertainty into confidence, to transform fuzziness into 
facts. These statistics don't even need to be particularly good 
numbers. We seem to believe that any number is better than no 
number, and we sometimes seize upon whatever figures are 
available to reduce our confusion. The problem is that a cer- 
tainty inspired by magical numbers may in fact be a poor guide 
for making decisions about the real world. 

This chapter examines types of numbers that, at least some- 
times, take on magical properties. These examples can help us 
understand the nature of magical numbers. We begin with a de- 
cision that confronts many families. 


Choosing a college is an anxiety-provoking process. The cost of 
a four-year undergraduate education is substantial and can be 
counted on to rise each year. Although the same can be said 
about the cost of a new car, customers who walk into an auto 
dealership with enough money are rarely turned away, whereas 
most applicants to elite colleges are denied admission. This un- 
certainty — will I get in? — leads students to apply to more than 

one institution, in an attempt to ensure that they are accepted 
somewhere. Applicants granted admission by more than one 
college are able to choose among these offers. 

But there are thousands of colleges out there, and the applica- 
tion process itself costs money. To which schools should students 
apply? A small industry has emerged offering guides to selecting 
colleges. Especially prominent is the newsmagazine US . News & 
World Report, which each fall publishes an annual guide for 
prospective students that ranks colleges based on statistical infor- 
mation. This issue, which sells far more copies than the maga- 
zine's regular weekly issues, is known among college admissions 
officers as the "swimsuit issue." The guide ranks colleges within 
categories ("Best National Universities -Doctoral," for instance), 
based on numeric scores, on seemingly objective criteria. 

Now stop and ask yourself what criteria someone would use 
to choose a college. How about quality of education? All things 
considered, a high-quality education ought to be more desirable 
than one of lesser quality. But it is very difficult to define quality 
of education, let alone measure it and then rank colleges by this 
measure. In fact, quality of education is likely to depend on all 
sorts of hard-to-predict things. We might suspect that this very 
year we can find students at every single college in the nation 
who are benefiting greatly, who are getting what are, for their 
purposes, high-quality educations, just as we can also find stu- 
dents on every single campus who are having rotten experiences 
and getting lousy educations. But, having said that, how can we 
hope to convert these experiences into numbers? Recall chapter 
is discussion of the difficulties with counting the incalculable. 

US \ News resolves this dilemma by ranking colleges accord- 
ing to criteria that are easy to quantify. This is a common solu- 

tion to this sort of problem. Colleges themselves, for example, 
want to promote professors who are good teachers and scholars, 
but it is very difficult to measure the quality of either teaching 
or scholarship. To make these decisions, most colleges rely heav- 
ily on criteria that produce numbers — scores on the teaching 
evaluations completed by students or the number of publica- 
tions a professor has written — even though everyone involved 
acknowledges that teaching evaluation scores and numbers of 
publications are only loosely related to faculty quality. These 
imperfect measures are at least numeric and therefore allow fac- 
ulty to be ranked: Professor A has better teaching evaluation 
scores than Professor B, and Professor X has more publications 
than Professor Y. Numbers seem objective; what we can express 
as a number often becomes the decisive measure, simply because 
the absence of numbers makes other criteria seem too arbitrary. 

What sorts of numbers can C/.5. News find for ranking col- 
leges? The magazine uses a complicated formula to create its 
rankings, but, for our purposes, we can focus on three sorts of 
figures incorporated in the formula. The first concerns the qual- 
ifications of the students the colleges admit. Because the maga- 
zine looks for indicators that can be reduced to numbers that 
are available from every campus, two measures emerge: scores 
on college entrance exams, such as the Scholastic Aptitude Test 
(SAT), taken by prospective students; and students’ high school 
class rank. The assumption is that colleges that admit better stu- 
dents — that is, students with better numbers (higher test scores 
and class standings) — deserve higher rankings. Once again, a 
qualitative concept is measured by the available quantitative 

The second set of measures concerns the college admission 

process itself. Here it helps to think of three stages: first, 
prospective students apply to a college; second, the college ad- 
mits some of those applicants (tells them that they are welcome 
to enter the college as students); and third, some of those admit- 
ted choose to attend that college. The number of students who 
decide to attend is important because colleges plan their budgets 
by assuming that they will have a certain number of students on 
campus in the fall. If too few students show up, the college will 
bring in less income than planned and will be forced to cut 
back; if too many students arrive, the college may not have 
enough professors, dorm rooms, equipment, and so on to ac- 
commodate them all. Because many students are admitted to 
more than one college, every college knows that some propor- 
tion of the applicants it admits will turn down its offer in favor 
of other institutions. Therefore, in order to be confident that 
enough people will show up next fall, colleges must admit more 
students than they can actually handle. 

The three stages produce three numbers: the number of ap- 
plications, the number admitted, and the number who accept 
admission. When we divide the second number by the first, we 
get the proportion of applicants who are admitted (called the 
admission rate). Dividing the third number by the second gives 
us the proportion of admitted students who accept the invita- 
tion to attend (called the yield rate). US . News uses the admis- 
sion and yield rates to rank colleges. The magazine assumes 
that a college that admits only a small proportion of those who 
apply (that is, it has a low admission rate) is choosy; it takes only 
the best students. And, if a large proportion of those admitted 
choose to attend that college (that is, it has a high yield rate), 
those choices indicate that students view the college as desirable. 

Once upon a time, admission and yield rates were internal 
figures, used by a college's administrators to plan. If you assume, 
for example, that this year's yield rate will be about the same as 
last year's, you have a reasonable idea of how many applicants 
you should admit in order to get the number of first-year stu- 
dents you want to arrive on campus. But now, thanks to US, 
News and the rest of the college admissions guidebook industry, 
these figures are not just public; they are also seen as a reason- 
able basis for comparing the quality of colleges and are part of 
the formulas used to calculate rankings. 

Such emphasis leads colleges to try to boost their rankings by 
improving the numbers that US. News uses in its calculations. 
One way to do this is to attract more applications — even if 
you're already receiving plenty of good applications, increasing 
the total number (while accepting the same number of students) 
allows you to report a lower admissions rate, thereby making 
your college seem more selective. Similarly, one of the reasons 
colleges like early-decision programs is that they attract appli- 
cations from students who are more likely to accept an offer of 
admission; increasing the number of these students raises the 
yield rate and thereby enhances the ranking." 

The third element in the US. News formula for calculating 
rankings is actually the most important: peer assessment, which 
counts for 40 percent of a college's score. 3 The magazine sends 
ballots to two officials at each college, who are asked to assign 
numeric scores to other institutions around the country. (On my 
campus of the University of Delaware, these ballots go to the di- 
rector of public relations and the associate provost for enroll- 
ment management, who oversees the admissions process.) Right 
away, questions arise. What qualifies these officials to assess the 

quality of other colleges? Why not send ballots to people more 
directly involved in educating students? What possible basis can 
these raters have for evaluating institutions they have probably 
never seen? Shrewd institutions now engage in direct market- 
ing — to these voters. In the weeks before the rating sheets ar- 
rive from US . News, the officials who will be casting ballots 
begin to receive advertising — glossy fact books sent out by var- 
ious colleges, each extolling the virtues of its campus. Presum- 
ably an effective campaign will result in higher scores from the 
raters, which will lead to higher rankings. 

Colleges tend to focus on these three elements in the US \ 
News formula because they are relatively easy to change. The 
formula also incorporates several other factors, such as the 
number of faculty members, spending per student, and the 
graduation rate, but these are hard to alter because it would be 
either too expensive or too difficult to change them substantial- 
ly from year to year. In contrast, encouraging more admissions 
and advertising the virtues of your campus to those who will 
cast ballots are relatively inexpensive tactics that might produce 
quick, favorable shifts in scores. 

None of these manipulations, of course, has anything to do 
with the quality of education a college offers. Yet they are im- 
portant, because year-to-year fluctuations in a college's ranking 
in the swimsuit issue can affect prospective students' application 
decisions. This is true even though shifts in the rankings are far 
more likely to reflect changes in how a college's admissions 
office conducts its business or how well the institution promotes 
itself to those who fill out the peer rating forms than anything 
that occurs in its classrooms. 

This example reveals how magical numbers work. Magical 

numbers help to resolve uncertainty. In this case, prospective 
students and their parents who want to make wise college deci- 
sions are confronted with a bewildering array of choices. The 
US . News rankings seem to offer an objective basis for making 
decisions: the swimsuit issue translates educational quality into 
a formula composed of quantifiable elements, and this formula 
produces numeric scores that allow us to rank colleges. Back- 
stage, some colleges may be working to improve their rankings 
not by actually improving education on their campuses, but by 
soliciting more applications or touting themselves to the public 
relations officers on other campuses. This activity remains hid- 
den, however. And those who want to place their faith in the 
swimsuit issue can take comfort in the belief that their decisions 
are rooted in nice, apparently solid statistics. 


Presumably, college rankings work their magic on individual 
students and their families. Many prospective students no 
doubt ignore these guidebooks, and, even among those who 
consult them, few are likely to make their college choices strict- 
ly on the basis of these rankings. The importance of the guide- 
books' statistics — the degree to which they seem to exert mag- 
ical power — varies among individuals. In contrast, other num- 
bers have greater influence; they may affect many people, more 
or less simultaneously, within particular organizations and 

While individuals sometimes turn to numbers to resolve un- 
certainty, most large organizations depend on statistics just to 
manage their day-to-day operations.' Organizations need num- 

bers to assess how well things are going. Businesses need to cal- 
culate costs and sales, profits and losses, while government and 
other nonprofit agencies have their own budgets and schedules. 
Organizations generate progress reports, efficiency reports, 
evaluations, assessments, and all manner of other number- 
crunching documents. The larger the organization, the more 
difficult it is to keep track of everything that is happening, and 
the more its managers and other members will depend on num- 
bers to summarize and clarify the complexity and to help them 
evaluate their own and others' performance. These figures con- 
dense reality into apparently straightforward measures; they 
provide the basis for the organization's decision-making. Still, 
the underlying process is not that different from bewildered 
high school students turning to college rankings: ambiguity and 
uncertainty encourage organizations to use statistics to simplify 
complexity. And, to the degree that these numbers become key 
to understanding and interpreting what is happening within 
the organization, the figures take on magical qualities. 

Whenever numbers are consequential, whenever people take 
them seriously and use them as a basis for decisions and actions, 
someone has a stake in those numbers. People who make deci- 
sions on the basis of statistics provided by others need to believe 
that those figures are correct, accurate, and valid — and they 
may try to ensure that they're given good data. In turn, the peo- 
ple who are affected by those decisions prefer numbers that lead 
to favorable outcomes; statistics that encourage your boss to in- 
crease your budget are clearly preferable to figures that might 
cause your boss to fire you. People care about numbers, and the 
more magical the number — the more it is treated as significant 
and meaningful, as the basis for decision-making — the more 

they are likely to care. And, since all numbers are produced by 
someone counting something, there are sure to be efforts to 
influence the production of — the counting that leads to — mag- 
ical numbers. We have already seen one such example: the var- 
ious attempts by colleges to raise their rankings in the US, News 
swimsuit issue. Analogous moves occur in most organizations. 

Organizational numbers take two principal forms: some are 
for internal use, while others are intended for external purpos- 
es. Internally, subordinates such as managers of particular de- 
partments might be required to report figures on expenditures 
or productivity to their bosses, who use these numbers to decide 
which units deserve more support or need closer supervision. 
Inevitably, complexity — all the factors that affect everything 
that is happening within the organization — gets condensed 
into a few numeric measures. But what is measured? When is it 
measured? How is it measured? The answers to such questions 
reflect choices about what counts within that organization. 
When a boss requires subordinates to report certain numbers, 
the assumption is that those figures can provide a picture of 
what's important. 

Requirements to report statistics to others within the organ- 
ization set the stage for bureaucratic "numbers games." Obvi- 
ously, a magical number that works in one's favor is a good 
number; subordinates have every reason to cooperate in pro- 
ducing such statistics. But if numbers imply that a unit has 
problems, it might be possible to minimize their impact. A 
canny subordinate might be able to manipulate the figures in a 
report in order to convey the best possible impression, perhaps 
even to suggest that this unit is doing a particularly good job, 
that it is more efficient, more productive, more deserving of re- 

ward than rival units. Alternatively, when the requested figures 
can't be massaged to provide a favorable picture, an experienced 
subordinate might argue that the measures are imperfect, that 
they fail to assess what is really important or to recognize what 
the unit does well, that these data are meaningless, and that al- 
ternative measures are in order. 

In turn, shrewd supervisors will be aware of their subordi- 
nates' interest in putting the best face on things, and they will 
try to ensure that the numbers they receive are accurate. In cases 
when suspect numbers are reported, supervisors might demand 
additional reports using new measures, or they might insist on 
specific, standardized methods of measuring and reporting. 
These new demands then invite subordinates to consider how 
they might also turn these new numbers to advantage. When 
supervisors fail to exert such control, the organization can be 
plagued by false figures. For example, the former Soviet Union's 
statistics on agricultural production — generated by subordi- 
nates more frightened by the penalties for reporting poor har- 
vests than by concern that their false reports might be discov- 
ered — stand as a monument to this sort of internal deception? 

Other numbers have external audiences; they are seen — and 
treated as meaningful — by people outside the organization. 
Investors, for example, use the figures in corporate financial re- 
ports to decide whether firms are attractive investments. When 
an organization is aware that outsiders will be examining the 
numbers it produces, its members may work to shape the num- 
bers in order to convey the desired impression to that audience. 
Once again, a "numbers game" is being played, only now not all 
the players are within the organization. Thus, some critics 
argue that because contemporary investors pay particular atten- 

tion to corporate quarterly earnings — "the Number" — corpo- 
rations now favor business policies and accounting practices 
that can generate favorable earnings that match or exceed mar- 
ket expectations, even if different actions might be in the firm's 
(and investors') long-run interest . 6 Similarly, police departments 
sometimes classify crimes in ways that minimize the crime rates 
in their cities, thereby implying that the police are doing an 
effective job . 7 Whenever outsiders are known to use magical 
numbers to assess organizational performance, the organization 
has opportunities to affect that assessment. (Remember those 
colleges trying to enhance their guidebook rankings.) 

In turn, knowing that an organization may manipulate its 
statistics, outsiders can try to gain a measure of control over the 
numbers. They might insist that the organization report certain 
information in certain ways. For example, when the FBI asks 
police departments to fill out the Uniform Crime Report forms 
that serve as the basis for calculating crime rates, the bureau 
gives detailed instructions for what to count and how to count 
it. Similarly, the U.S. Securities and Exchange Commission 
specifies a general format for corporate financial reporting. 
These are efforts to make reports from different organizations 
comparable. The outsiders may even try to establish and enforce 
penalties for those who disseminate incorrect numbers. Such 
measures can discourage deceptive reporting, but, as the Enron 
scandal reminds us, they cannot ensure that the reported figures 
will be accurate . 8 

The point is that organizations need statistics to operate, both 
to provide a basis for their internal decisions and as a means of 
summarizing their activities to outsiders. But to the degree that 
people either inside or outside the organization take those fig- 

ures seriously and use them as a basis for decisions — that is, the 
more magical the numbers are — the more the organization's 
members have a stake in shaping the statistics to match their 
own interests. It would be naive to imagine that statistics re- 
ported by organizations simply mirror reality, that they reflect 
the simple, whole truth. We must acknowledge that there are 
trade-offs. Organizational numbers always condense complexi- 
ty, which has both benefits and costs: such numbers allow us to 
summarize, to clarify, to grasp the big picture; but these sum- 
maries inevitably simplify, as people choose what to count and 
how to count it. The more consequential (magical) the numbers 
are, the more likely people are to think carefully about those 
choices and work to make the numbers convey their side of the 
story, and the less confidence we can have in the figures as a 
straightforward reflection of reality. 

This is the paradox of magical numbers: we need them, and 
we need to be able to trust them; yet the greater our need, the 
more likely that the figures will be distorted, and the more care 
we must take when examining them. Before we can rely on sta- 
tistics, we need to ask who counted what, and how and why 
they counted it — because, as our next example shows, when 
magical numbers become the focus of widespread attention, the 
potential for confusion is very great. 


Anxiety about the quality of American education grew to re- 
markable levels during the last decades of the twentieth centu- 
ry. This might seem curious. After all, Americans' average years 
of schooling increased dramatically throughout the century. In 

1900, only about 6 percent of American seventeen-year-olds 
graduated from high school, 9 By the century's end, most Ameri- 
cans were continuing their education beyond high school, and 
about a third of those in recent age cohorts completed bachelor's 
degrees. The United States now has one of the largest percent- 
ages of highly educated citizens in the world. 10 Other statistics, 
however, were troubling. Studies found that Americans stu- 
dents often scored less well on comparative tests than students 
in other countries, particularly in math and science, subjects in 
which cultural differences should have only a minimal impact 
(since the answers are either right or wrong). This comparison 
indicated that American students weren't learning as much or 
as well as their counterparts elsewhere. And scores on the SAT, 
the principal college admissions test, dropped from the mid- 
1960s through about 1980 , which suggested that the perfor- 
mance of American students might actually be getting worse. 
(Since 1980 , SAT math scores have largely recovered, although 
verbal scores have remained low.)” 

This evidence raised doubts about the quality of American 
education and student accomplishment, and critics expressed 
concerns about what this might mean for the country's future. 
Perhaps it was once possible to drop out of school and still make 
a reasonable living, but the modern job market requires more 
education — it offers fewer jobs that demand strong backs and 
more that need nimble brains. Today's drop-out seems to be 
risking a lifetime of marginal poverty. Moreover, a country that 
fails to maximize its citizens' education risks falling behind 
other nations that do a better job of educating their young. Nor, 
these critics warned, should Americans find comfort in the 
higher rates of school completion; graduating larger numbers of 

less able students is simply proof that schools have abandoned 
academic standards. These critics offered a nostalgic vision of 
the educational past: in the good old days, students worked 
hard; they really learned their lessons; they were dedicated, de- 
termined. But these kids today! They don't know things, they 
don't care, they don't read, they watch television, and their 
music — if you can call it music. . . . Inevitably, the critics began 
to sound like their parents. 

Nostalgia offers a faulty lens for viewing change. Each gen- 
eration's educational critics tend to be people who themselves 
did pretty well in school, at least well enough to acquire the cre- 
dentials to become critics. They remember themselves and their 
friends as being fairly good students, and they often forget their 
classmates who did less well or who may have left school. Con- 
trasting the critics' memories with today's students does not 
necessarily compare apples with apples. Even changes in stan- 
dardized test scores may prove tricky to evaluate. If we assume 
that, in general, the more able students stay in school the 
longest, then, as the share of young people who remain in high 
school or enter college increases, the average abilities of high 
school graduates or college students might decline because more 
lower-performing students are continuing to pursue education. 
Thus, measuring educational achievement across time may well 
compare scores from rather different populations of students. 

Still, criticizing schools appeals to all sorts of critics. Con- 
servatives can blame poor performance on schools having drift- 
ed away from a traditional academic curriculum and strict dis- 
cipline, and they call for a return to these fundamental princi- 
ples. Liberals can argue that schools are failing to reach students 
who are somehow disadvantaged and that the curriculum needs 

to be modified to educate those most vulnerable students. What- 
ever a critic’s particular agenda, most agree that something 
must be done, although the critics probably won't agree on just 
what that something should be. And a society that preaches per- 
fectionism — "we will leave no child behind" — seems particu- 
larly likely to see schools as falling short and criticisms of edu- 
cation as well founded. 

Recently, this anxiety about education has led to the wide- 
spread adoption of standardized educational testing as a means 
of holding schools accountable. The states and the federal gov- 
ernment require that all public school students be given stan- 
dardized tests and that the results — particularly the average 
scores at different schools — be made public. These tests have 
various consequences. For individual students, poor test scores 
may lead to mandatory summer school to help them catch up; in 
some school districts, students who complete the required cours- 
es but who cannot achieve some minimum test scores may re- 
ceive a lower grade of high school diploma than their higher- 
scoring classmates. Teachers also face consequences. Some advo- 
cate that teachers whose students perform better on the tests 
should receive larger "merit" salary raises than colleagues whose 
students do not do as well. In addition, schools are singled out. 
Newspapers report the test results by school, implying that some 
schools are doing a better job than others; in some states, schools 
that show unusually large improvements in test scores receive 
awards. The implications reverberate outside education: realtors 
find that being located in a high-scoring school district has be- 
come a selling point for houses. The test scores, in short, have be- 
come an especially vivid example of magical numbers. 

For educational testing to have serious consequences for stu- 

dents, teachers, and schools, we must make certain assumptions 
about what the tests measure. Most obviously, we must assume 
that the tests provide a valid measure of students’ learning and 
abilities, that students who receive higher math scores actually 
have learned more math. But we must also assume that the 
teaching that occurs in schoolrooms is the key to this learning. 
At first, this might seem beyond dispute — "Isn't learning exact- 
ly why we send students to school — and isn't teaching exactly 
what schools are supposed to do?" 

But note the familiar role of social class in schooling: in gen- 
eral, upper-middle-class (disproportionately white) students 
tend to do better in school than lower-class (disproportionately 
black or Latino) students. The causes for this pattern are hotly 
debated. Various explanations emphasize differences in the stu- 
dents (for instance, arguments that intelligence is determined 
partly by genetics), differences in the students' social circum- 
stances (for example, whether family, friends, and other people 
in the students' lives value and support education), and differ- 
ences in schooling (such as whether upper-middle- class chil- 
dren attend schools with better teachers, smaller classes, nicer 
facilities, and a variety of other advantages). Different explana- 
tions carry varying implications for testing policies. Thus, if we 
assume that what happens in school is the principal factor in de- 
termining how much students learn, then test scores might be a 
good index of school performance. But if we assume that stu- 
dents' social circumstances have powerful effects on shaping 
learning, then test scores may ultimately measure little more 
than the students' social class. 

This is one reason debates over testing policies have become 
so acrimonious. Advocates of testing argue that schools and 

teachers ought to be doing a much better job and that they must 
be held accountable by using students' test scores as the measure 
of the educators' performance. Presumably, if educators do the 
job they are paid to do, their students will pass the tests. If stu- 
dents at some schools perform poorly on the tests, they can be 
required to take summer school, they may not qualify for aca- 
demic diplomas, their teachers should receive lower merit rais- 
es, and so on. In short, the scores will have serious consequences. 

As these consequences have become apparent, critics of test- 
ing have become more vocal. Not surprisingly, many teachers 
and principals oppose testing systems that penalize educators 
for students' poor performance. Teachers' unions argue that 
teachers are being blamed for the social circumstances of their 
students: "Of course children who come from upper-middle- 
class homes filled with books, who have two educated parents 
who emphasize the importance of education, and who benefit 
from other advantages do well in school. Lower-class children 
who lack those advantages can be expected to have more trou- 
ble learning, and we should not blame the teachers for things 
they can't control." Other critics argue that even good test scores 
may be an illusion, because high- stakes testing will lead schools 
to "teach the test," that is, to drill their students in the sorts of 
questions that appear on the tests, while ignoring other, perhaps 
more important, forms of learning. Such critiques are precisely 
the sorts of reactions we ought to expect from those whose per- 
formance is being assessed when large institutions adopt magi- 
cal numbers: teachers (who are being evaluated by their stu- 
dents' scores) argue that tests cannot possibly accurately meas- 
ure whether teachers are doing a good job, while nonteachers 
suspect that teachers may alter their instruction in ways that 

maximize their students' scores but diminish the actual quality 
of the teaching. There are even reports of teachers helping stu- 
dents cheat in order to improve their scores.'" 

We also encounter other problems with the way scores are put 
to use. For example, some states award special recognition to 
schools that show marked improvement in year-to-year test 
scores, a practice that is probably misguided. Research has 
shown that the schools with substantial year-to-year shifts in 
scores tend to have fewer students , 13 which suggests that the 
numbers taken to measure excellence in teaching may be noth- 
ing more than statistical artifacts. Imagine a small elementary 
school with a single classroom for each grade. The year-to-year 
scores in such a school are relatively volatile; if this year's class 
contains just a couple of very good students, this year's third- 
graders may score markedly higher than last year's class filled 
with ordinary students, even if the teacher taught exactly the 
same lessons both years. Such year-to-year variation is less likely 
in a larger school with, say, five rooms of third- graders; there, a 
couple of bright students will have less effect on the school's per- 
formance, and test scores are likely to remain fairly stable. Even 
if students score randomly on tests, small schools are much more 
likely to have their scores increase — and decrease — from year to 
year than large schools. But, of course, when test scores are treat- 
ed as magical numbers — the definitive measure of how well 
schools and teachers and students are doing — the possibility that 
chance might play a role in shaping scores disappears from poli- 
cy discussions. 

Educational testing, with its promise of bringing schools 
under control, is in vogue and promises to remain there for a 
while — at least until this policy's limitations become more ap- 

parent. It offers a clear example of the power of magical num- 
bers, and it ought to serve as a caution for other would-be nu- 
meric reformers. 


Another contentious issue in recent years has been the practice 
of racial profiling by police. We need to begin by recognizing 
that different people use the term metal profiling to refer to very 
different things. Here, I will restrict my discussion to claims 
about police stopping cars partly on the basis of the driver's 
race. Many police officers argue, and the courts have generally 
agreed, that race may sometimes be considered a relevant char- 
acteristic — not the sole reason, but one of several — in deciding 
to stop a car. Suppose, for example, that police have reason to 
believe that drugs are being transported along a particular 
route, in particular sorts of vehicles, by couriers for an African 
American criminal network; under these circumstances, police 
might decide to stop a suspicious vehicle in part because its driv- 
er is black. This is how defenders of racial profiling tend to de- 
scribe the policy. 

In contrast, critics of racial profiling talk about being pulled 
over for "DWB" ("driving while black"). Many African Ameri- 
cans believe that they are far more likely than white drivers to 
be stopped by police, because police suspect that blacks are more 
likely to be involved in criminal activities. In such cases, a dri- 
ver's race may be the sole basis for stopping a vehicle." In this 
view, racial profiling is a racist practice. Some critics argue that 
race should never be a consideration in stopping a vehicle. 

Almost as soon as racial profiling emerged as a visible political 

issue, people began calling for the collection of statistics that 
could determine, once and for all, the existence and extent of the 
practice. That is, they demanded a magical number, a measure of 
racial profiling. Collecting statistics has become a popular com- 
promise measure in contemporary politics; for instance, the first 
federal law concerning hate crimes required the FBI to begin 
counting hate crimes in order to measure the scope of the prob- 
lem. Such compromises imply that statistics can magically re- 
solve disputes. Statistics are viewed as factual, as offering a clear, 
unbiased portrait of police practices, hate crimes, or whatever 
else is at issue. It is difficult to oppose collecting such statistics be- 
cause data collection is assumed to be nothing more than deter- 
mining the facts. Besides, the participants in a debate may all be- 
lieve that the statistics will support their position: in the case of 
racial profiling, the police may anticipate that such statistics will 
reveal that they behave responsibly, whereas their critics may as- 
sume that the numbers will expose discriminatory practices. 

The problem is that measuring racial profiling is likely to be 
much trickier than we might think.” The simplest studies of 
racial profiling compare the race of drivers stopped to the racial 
composition of the area's population. Suppose that 20 percent of 
drivers stopped by a town's traffic officers are black. Before we 
can interpret that finding, we need to know something about 
the population of drivers on the road. Are 10 percent of the 
area's drivers black (which would suggest that African Ameri- 
cans are stopped far more often than might be expected)? Are 
20 percent of the drivers black (which would suggest that the 
proportion of African Americans stopped is about what we 
would expect)? Or are 30 percent of the area's drivers black 
(which would suggest that African Americans are stopped less 

often than other drivers)? Making such comparisons is simpler 
in theory than in practice, however. 

The key issue is how to identify the population of drivers 
that should be used as the basis for comparison. One criminolo- 
gist calls this problem "searching for the denominator ,” 16 The 
easiest basis for comparison is the racial composition of a town's 
population (available from census statistics). But notice that not 
everyone in a town's population drives; presumably, we ought to 
adjust our population estimate by trying to determine the racial 
composition of the town's licensed drivers. But not all licensed 
drivers drive the same number of miles — and we might assume 
that the more one drives, the greater the risk of being stopped 
by the police. 

In addition, if some roads are driven mostly by locals, the 
drivers presumably reflect the community's population. But 
other roads, such as interstate highways, carry a large propor- 
tion of drivers from elsewhere, who will not reflect the local 
population. The driving population probably changes from 
daytime to nighttime, and weekday to weekend, and we should 
not be surprised to find that police decisions to stop drivers may 
depend on time of day. For example, officers might be more 
likely to stop someone for reckless driving late at night, on the 
grounds that late-night drivers might be intoxicated. In short, 
getting people to agree to gather data on the racial distribution 
of drivers who get stopped by police is only part of the problem; 
we also need to agree on a basis for comparison. 

A somewhat more sophisticated approach is to try to meas- 
ure the race of traffic violators. In one early study, researchers 
drove at the speed limit down a stretch of interstate highway in 
Maryland and observed all the cars that passed them (which had 

to be speeding and therefore were theoretically eligible to be 
stopped by the state troopers who patrolled the road). The study 
found that 18 percent of the speeding drivers appeared to be 
black, whereas 28 percent of the drivers stopped by the Mary- 
land state police were black. 

While the results of this study were certainly suggestive, it is 
not difficult to identify its flaws. Everyday experience tells us 
that a substantial proportion of drivers exceed the speed limit, 
but that police ordinarily will not stop a driver going slightly — 
say, up to ten miles per hour — above the limit. Therefore, a 
study that treats all drivers who exceed the limit as eligible to be 
stopped may not have identified the relevant population. If, for 
example, whites are relatively more likely to drive just a few 
miles above the limit, while blacks tend to drive faster than that, 
then the proportion of those driving fast enough to attract po- 
lice attention who are black might be greater than the percent- 
age of African Americans among those drivers who exceed the 
speed limit. 

All manner of other complexities suggest themselves. Speed- 
ing in and of itself may not be what leads police to stop cars. 
Perhaps they are equally — or more — interested in reckless 
driving. Perhaps they focus on older cars, which might be more 
likely to have visibly faulty equipment. If African Americans 
drive older cars, or more often drive recklessly, this might help 
account for them being stopped more often. Or perhaps there 
are demographic differences between drivers of different races. 
We know that young drivers get into more accidents. If the 
population of black drivers contains a larger proportion of 
young drivers, we might expect them to attract a disproportion- 
ate amount of attention from the authorities. 

The point is not that any of these explanations is necessarily 
true. Rather, it is that using the race of drivers who pass a re- 
searcher's car that is moving at the speed limit is an imperfect 
way to identify drivers whom police might decide to stop. The 
Maryland study's statistical findings are suggestive, but they are 
hardly ironclad proof of the extent of racial profiling. 

In short, measuring racial profiling is not the simple, 
straightforward matter that it might seem. However data on 
racial profiling are collected, some will argue that the resulting 
statistics are illegitimate. The call to gather data seems based on 
the belief that these statistics will be generally accepted as mag- 
ical numbers, but it is unlikely that everyone will grant these 
figures that sort of authority. 

These problems do not necessarily mean, however, that data 
collection wouldn't be worthwhile. We might suspect that police 
departments that collect data on the race of the drivers stopped 
by their officers might find the information useful. The discov- 
ery, for example, that some officers — or even one particular 
officer — stop far larger percentages of African American drivers 
than other officers patrolling the same streets would seem to 
raise legitimate issues. The officers in question might be asked to 
explain why their pattern of stops differs from those of their col- 
leagues. Even the knowledge that information is being collected, 
that a record of one's performance will be reviewed, may en- 
courage police officers to evaluate their own actions, to make 
sure that their traffic stops are appropriate and justifiable. (Some 
critics warn of another outcome: officers manipulating their 
records to obscure evidence of racially based actions. Once more, 
we see how an organization's decision to keep statistical records 
might lead its members to try and shape the resulting numbers.) 

Collecting and examining data on the race of drivers stopped 
may well lead to desirable outcomes. But such data should not 
be understood as somehow providing a precise, perfect measure 
of the extent of racial profiling. Every attempt to measure racial 
profiling will require making choices, choices that someone 
may question. The resulting numbers may have their uses, but 
they also will have their flaws, and people might have reason to 
question their magical status. 


The examples in this chapter illustrate a dilemma. We live in a 
complicated world, and we need statistics to help make the 
complexity understandable. We tend, then, to seize upon what- 
ever numbers are available, to treat them as factual, accurate 
distillations of reality — in other words, we treat them as if they 
have a magical power to summarize and clarify, to provide a 
firm basis for decisions. But as soon as people become aware that 
someone has begun to treat a number as magical, the "number 
games" begin; folks try to manipulate the number so that the 
magic can work in their favor. 

Once we understand this process, we should appreciate two 
reasons why we need to handle magical numbers with special 
care. The first, of course, is that we must consider the choices 
that underpin these statistics. People distill complexity into sim- 
plicity by making choices, by highlighting some features and 
dropping others from consideration. Such choices are both in- 
evitable and consequential. Yet once we are given a number, we 
often forget to consider how those choices shaped the outcome. 
Remembering this process is essential if we are to avoid being 

taken in by magical numbers. The second concern is that we 
need to be especially alert to the possibility that people with a 
stake in the outcome may have manipulated these figures. The 
more magical the number, the more likely it is that someone 
affected by it will try to play a numbers game, and the more im- 
portant it is to question how and why people created the figure 
and how its magical status affects the ways people count. 


I t is no trick to spot controversies about statistics. Argu- 
ments over numbers make the news. Have Hispanics be- 
come the nation's largest ethnic minority? Should federal 
guidelines for acceptable levels of arsenic in drinking 
be modified? Is hormone replacement therapy beneficial or 
dangerous? Such questions highlight debates about data. 

The widespread assumption that statistics can reduce com- 
plexity to summaries of simple facts is more than just a way of 
attributingmagical power to numbers. It is also a way to win ar- 
guments. In debates over social and political questions, people 
sometimes present statistics as though they are rhetorical trump 
cards, facts that can overwhelm any opposition. Because figures 
are considered such powerful evidence, they often cannot be ig- 
nored but must be challenged, either with questions about their 
accuracy or with rival numbers. As a result, people who intro- 

duce statistics in order to win debates may find themselves ar- 
guing about numbers. 

Not all statistics inspire strong opposition. Some advocates 
address matters of consensus. Child pornography, say, has few 
defenders. Statistical claims about such topics of consensus can 
get a free ride; often, no one inspects them closely. But other so- 
cial issues become matters of bitter debate because they invoke 
competing ideologies or interests. And, where there is a clear 
basis for opposition, statistics offered by one side regularly draw 
critiques from the other. 

These statistical controversies — what I've called stat wars — 
take different forms.' The simplest disputes concern the accuracy 
of a single number. A figure is brought to people's attention, only 
to have its accuracy challenged for some reason ("is that really 
the correct number of alcohol-related traffic deaths?"). Often, the 
issue is whether the people counting have done a careful and 
complete job, whether their definitions or methods might have 
led to a number that is too high or too low. In some cases, as 
when statistics are merely estimates, the number can be easily 
called into question. It is common for a lone number to be ad- 
vanced — and challenged — because it serves as a kind of short- 
hand proof for some claim ("this problem needs to be treated se- 
riously, as evidenced by our large estimate for the number of 
cases"). To the degree that a number is central to the argument, 
opponents will challenge that figure, either by pointing to rea- 
sons to doubt the number or by countering the original estimate 
with one of their own. 

Debates over single numbers tend to occur early in the his- 
tory of public issues, when people are trying to draw attention 
to a social problem, before they have had time to collect a lot of 

information about the topic. One sign that an issue has ma- 
tured is a proliferation of statistics: more people start counting 
more of the problem's elements in more ways. A body of re- 
search studies may emerge; some topics may generate hun- 
dreds or thousands of numeric findings, with advocates sifting 
through them in a search for statistics that seem rhetorically 
powerful. As the pool of available statistics expands, so do op- 
portunities to locate figures that one can use to support differ- 
ent stances. Advocates who search a sufficiently large pool of 
data can probably come up with evidence to support whatever 
position they favor, but their opponents are also likely to find 
figures that they can use to make the opposite case. Soon, sta- 
tistics become weapons, rhetorical grenades lobbed at the op- 
ponents' positions. 

Those who already favor a particular position in one of these 
debates find comfort in their side's numbers, while the opposi- 
tion's figures strike them as dubious, perhaps even fraudulent. 
Those of us who don't have a stake in an issue — the uncom- 
mitted public is often the target audience for competing nu- 
meric claims — can become frustrated by the flow of apparently 
contradictory numbers. "Just tell us," we snarl, "which chemi- 
cals cause cancer." We don't want a bunch of contradictory sta- 
tistics — we want the simple facts. 

But facts are socially constructed. What we recognize as 
facts are products of people's efforts to make sense of the 
world, to assemble enough evidence to support a general 
agreement that something is true. I am not arguing that there 
is no real world against which we can check our facts — there 
is. We all know that when we hold a rock in front of us and let 
go, it will fall down. Insisting that it will remain suspended in 

space won't make that happen. Still, what knowledge is con- 
sidered factual varies from time to time and place to place: for 
example, the most authoritative explanations for the causes of 
disease vary from one society to the next and across historical 

What we deem factual depends on a combination of evi- 
dence and consensus. Evidence matters; claims that germs 
might cause disease received a huge boost when microscopes let 
people see microbes. But consensus is also necessary; it took 
time and considerable research before medical opinion came to 
a general agreement about the value of the germ theory. Over 
time, the boundaries of consensus expand, although areas of 
dispute may remain. When we grumble that news stories about 
what is or isn't a cancer threat seem to change from week to 
week, we are complaining about a lack of consensus — which, 
in turn, probably reflects available evidence that is weak or 
deemed inconsistent. 

There is an important point here. Debates about what is true 
tend to polarize around two weak positions. At one pole are the 
relativists, those postmodern theoreticians who imply that real- 
ity is up for grabs, that we can't really know anything, that we 
should be open to every alternative perspective andsuspicious of 
any purported authority. The extreme version of this position 
justifies all manner of paranormal beliefs, conspiracy theories, 
and other ideas grounded in little or no evidence. The other 
pole is the realm of the absolutists, who insist that facts are facts 
and who have no patience with challenges to authoritative 

This book argues for a position somewhere between these 
extremes. We are social beings. Everything we know about the 

world, every number and, for that matter, every word we use 
while thinking, is shaped by our social life. Anyone who has 
seen an infant grow into a child knows that we all had to learn 
language — and, in the process, we also learned our culture's 
way of dividing the world into categories. The great contribu- 
tion of classical anthropology was to demonstrate cultural di- 
versity, the many different ways people could make sense of 
their worlds. Every culture has ideas about why people get sick, 
expectations for how modestly young women ought to behave, 
and so on — and every culture believes that its ideas and expec- 
tations are right and true. To understand our world, we must 
recognize that all knowledge is filtered through peoples' cul- 
tures. In short, there has to be a place for relativism. 

On the other hand, science offers a particularly useful stan- 
dard for evaluating some sorts of knowledge about the world. 
Science is a process by which ideas are tested in ways that might 
disprove them; ideas that survive these tests are considered 
more likely to be true. Over time, this process produces knowl- 
edge in which we have great confidence. This process is not per- 
fectly smooth: findings may be initially accepted but later with- 
drawn when further tests call them into question; ideas may be 
ignored or rejected but later achieve acceptance; and so on. But 
these irregularities in assembling scientific knowledge should 
not be taken as evidence that the process doesn't work over the 
long run. 

I wrote the first draft of this paragraph on a computer, a ma- 
chine that is the product of centuries of gradually increasing 
scientific knowledge. I have great confidence that the machine 
will work, even though I must confess that I have only a prim- 
itive understanding of the scientific principles by which it oper- 

ates. Yet it would be silly for me to argue that the science be- 
hind that computer was essentially arbitrary, just one of many 
ways of thinking about the world, no better or worse than 
any other. The computer works. Vaccinations work. Scientific 
knowledge is not just one view among other, equally valid per- 
spectives. We can have great confidence in well-established 
scientific findings. In short, there has to be a place for authority 
grounded in evidence. 

Still, science cannot answer all questions. It can tell us how 
and why some people get sick (though it cannot, at this point, 
explain all illness). But it cannot tell us how modestly young 
women ought to behave; that is not a topic subject to scientific 
evaluation. The limitations of science pose a problem in our cul- 
ture, precisely because we have such high expectations for sci- 
ence. When we fall ill, we expect that a physician will be able to 
diagnose and treat what's wrong, and we become frustrated 
when this doesn't happen. We even use research documenting 
social patterns or assessing risks to recommend ways to behave. 
Our society treats data — statistics — as offering, if not complete 
answers, at least information relevant to devising the answers 
for many kinds of questions, including many that do not neces- 
sarily fall within the purview of science. 

When confronted with statistics, we need to avoid the poles 
of both extreme relativism and extreme absolutism. We need to 
remember that statistics are social products and that the process 
by which they are created inevitably shapes the resulting num- 
bers. But we must also appreciate that science offers ways of 
weighing the evidence, of assessing the accuracy of figures. 
These concerns become particularly important when statistics 
become the subject of disagreements. 


Jun\ science is a term, currently in vogue, used to dismiss find- 
ings as products of dubious research. Because science is consid- 
ered a source of authoritative knowledge in our culture, many 
people call themselves scientists as a way of legitimizing their 
views. Billing some set of claims as "scientific" is a modern way 
of claiming legitimacy and authority. Thus, some religious op- 
ponents of teaching evolution argue that they represent "cre- 
ation science," and they insist that the Biblical account of cre- 
ation ought to have equal footing with the explanations ad- 
vanced by physical and biological scientists for the origins of the 
universe, the Earth, and human life. Similarly, all manner of 
parapsychologists, psychic healers, and perpetual-motion advo- 
cates label their views "scientific."" 

But science is more than a name; it is an orientation toward 
evidence. Scientists must be prepared to test their ideas, and it 
must be possible for the tests to disconfirm those ideas. This is 
not quite the simple, pure process of hypothesis testing that 
junior high school textbooks describe. Scientists are people, 
and they may get caught up in their ideas, sometimes making 
excuses when those ideas fail in tests — something wasn't right 
with the test conditions, further tests are needed, and so on. 
We can point to the foibles of scientists who cling to their ideas 
in the face of challenging, even disconfirming evidence; focus- 
ing on such behavior allows us to draw a portrait of science 
that emphasizes its warts and flaws . 3 Some relativist critics 
argue that disagreements within science render it just one 
more viewpoint, no truer than any other. Perhaps one way out 
of this tangle is to recognize science as an ideal, but to ac- 

knowledge that individual scientists may fall short of this 

Nevertheless, over time, as the available evidence grows, sci- 
ence accumulates a body of knowledge in which we have great 
confidence, based on the reliability with which its predictions 
are confirmed. This scientific progress depends on a communi- 
ty that demands rigorous, continual self-examination, subject- 
ing ideas to tests that can determine whether the evidence sup- 
ports the ideas. Because every test has weaknesses, it is the cu- 
mulative application of multiple tests that provides the founda- 
tion for science’s eventual acceptance of only those ideas that 
hold up under the most vigorous examination. 

Single studies, then, can't do the job. Absolutely every study — 
every test, every piece of research — has limitations and flaws in 
its methods that make it a target for legitimate criticism. Studies 
should be replicated, and they should also inspire further research 
that uses different methods (with, presumably, different limita- 
tions and flaws). When replication and differing methodologies 
confirm the same result, confidence in that finding grows. The re- 
sults of a lone study, particularly if the research raises serious 
methodological concerns, should not, in most scientists' view, be 
treated as authoritative. Only time and further research can sort 
out the erroneous findings from the more reliable. 

Unfortunately, news coverage of scientific research tends to be 
less patient than the scientific community! The news media 
favor stories that seem novel, unexpected, dramatic. The most 
compelling scientific news story is about a sudden breakthrough, 
not a replication or a confirmation of an earlier finding using a 
different research design. Thus, the press prefers reporting ex- 
actly those research results that lack strong substantiation. A 

single study with a disturbing finding makes good news, and 
the media coverage is likely to downplay or even ignore the re- 
search's methodological limitations. As a consequence, we must 
approach press reports of research results with caution. This is 
particularly true given the efforts by some prestigious journals to 
promote their visibility by issuing press releases that highlight 
the most dramatic findings in articles they publish (as discussed 
in chapter 4 ). 

The pejorative label "junk science" typically implies a meth- 
odological critique, an argument that the research was designed 
or the data collected in ways that make it impossible to have 
confidence in the results. Often, it also implies that the research 
was guided by a particular agenda, shaping the findings to sup- 
port a specific position. The original usage of the term jun\ set - 
ence was to characterize expert witnesses' testimony in trials? 
Lawyers ask expert witnesses to testify in hopes that their ex- 
pertise will persuade judges and juries that particular argu- 
ments are factually true, supported by scientific research. When 
an expert witness is invited to testify (and is paid) by one side in 
a trial, it is reasonable to wonder whether that testimony will be 
complete, even-handed, and actually representative of scientific 

Consider, for example, the issue of "toxic" breast implants. 6 
In the late 1980s and early 1990s, the health risks of breast im- 
plants became a subject of considerable public concern: the 
Food and Drug Administration banned silicone-filledimplants; 
the issue received extensive media coverage; and a multibillion- 
dollar class-action lawsuit was filed. Critics of the implants were 
bolstered by various medical and scientific experts who present- 
ed evidence that a number of women who had implants experi- 

enced certain diseases. We see a familiar line of reasoning here: 
someone falls ill, tries to understand what caused the illness, re- 
calls some experience (such as having breast implants), and con- 
cludes that the experience must have caused the illness. The 
logic may seem perfectly compelling to the individual, but it 
cannot be considered scientific proof. 

Science demands, among other things, epidemiological sup- 
port. For example, we know that some people get sick, so we 
should expect some level of sickness among women who have 
breast implants simply because they are people. Therefore, the 
key question is whether women with breast implants are any 
more likely to fall ill than other, similar women who have not 
had implants. (Recall chapter j's discussion of risk: the usual 
standard for such comparisons is that the rate of illness should 
be at least 200 percent greater among women with breast im- 
plants than in the control group before we can conclude that 
implants probably cause disease.) In general, epidemiological 
studies did not show such higher rates of disease among women 
with implants. This evidence should have been viewed as a very 
serious challenge to claims that implants were harmful, but crit- 
ics of implants won the public relations battle (and many of the 
court cases), partly because the results of the epidemiological 
studies did not become known until very late in the issue's 

One problem with the notion of junk science is that the term 
has become politically loaded: conservatives often use it to dis- 
miss claims by environmentalists, consumer advocates, and 
other activists warning about dangers in contemporary society? 
In response, liberal critics argue that "the concept of junk sci- 
ence serves as a convenient way of reconciling. . . pro-corporate 

bias with pretensions of scientific superior ity.” s Each side argues 
that scientists on the other side are biased and cannot be trusted 
to design legitimate research. It is difficult for nonspecialists to 
assess these claims and counterclaims, if only because the differ- 
ences in research findings may derive from competing assump- 
tions, definitions, or methodological choices. For example, sci- 
entists working with environmentalists may define infrequent 
exposure to a very low concentration of a radioactive substance 
as a dangerous health risk, whereas scientists working for in- 
dustry may argue that more frequent exposure to higher con- 
centrations of the same substance does not pose an unacceptable 
risk , 9 Both sides may insist that theirs is the scientifically sound 
position, that their method of assessing risk is appropriate — 
leaving nonscientists frustrated by the need to weigh the claims 
of dueling experts. 

Although its links to particular ideological positions may 
make it impossible to rehabilitate the term jun\ science , a useful 
idea is lurking in this debate. Every piece of research contains 
limitations; researchers inevitably choose specific definitions, 
measures, designs, and analytic techniques. These choices are 
consequential; they shape every study's results. We can never 
have as much confidence in the results of any single study as we 
can in a body of research, in which the various researchers' 
choices help cancel out one another's limitations. Our 
confidence that smoking causes lung cancer is not founded on 
any single study, but on a large body of studies using different 
methods that — overall — link smoking with cancer. To be sure, 
some researchers have biases that lead them to design research 
in ways that may foster the results they favor; don't forget that 
the Tobacco Institute once sponsored research intended to gen- 

erate results suggesting that smoking was not especially harm- 
ful. But the real problem with much of what is called junk sci- 
ence is not so much the researchers’ motives or politics as it is 
the advocates' tendency to proclaim one or two preliminary 
studies as definitive. In such cases, the process of assembling 
scientific data gets short-circuited by political concerns. 

Debates over junk science have another notable feature. 
They tend to involve disagreements about notions of trade-offs 
and risks (raised in chapters i and 3) — will this chemical (med- 
ical procedure, hydroelectric project) cause unacceptable harm? 
Reasonable people might disagree about all sorts of issues here. 
How should we measure prospective harm? How should we 
weigh the harms (or costs) against the projected benefits (and 
how should we measure those)? While advocates may try to 
characterize such debates as contests between good and evil, the 
evaluation of scientific evidence is rarely so straightforward. 


Debates over social statistics rarely begin as disputes about a 
number. Rather, they almost always start as disagreements 
about the importance of a social issue or the solution to a social 
problem, with advocates proceeding to introduce numbers as 
ammunition to reinforce one position or another. Recent politi- 
cal discourse refers to spinning , the practice of offering the 
media an interpretation of events that coincides with one's own 
viewpoint, in hopes that the media will repeat — and possibly 
even endorse — that viewpoint. 10 Numbers can be subjects of 

Consider the conflicting interpretations offered when offi- 

cials announced that the 2000 census revealed that a growing 
proportion — about one-quarter — of households were com- 
posed of lone individuals.” For conservative, pro-family advo- 
cates, this statistic was further evidence of the decline of the tra- 
ditional American family, of the need for social policies to pro- 
mote families. But other, more liberal commentators interpret- 
ed the increase in single-person households in more positive 
terms: growing affluence and improved health meant that 
young people could afford to set up independent living arrange- 
ments, that individuals could end unsatisfactory marriages, and 
that the elderly could maintain their own households. Thus, 
one could read census statistics documenting the growth in 
single-person households as revealing either societal decay or 
improved living circumstances. Note that no one disputed the 
statistic's accuracy; people can acknowledge that a number is ba- 
sically correct without necessarily agreeing about what it means. 
The glass can be seen as half-full or half-empty — it just de- 
pends on the spin. 

The existence of well-articulated, competing ideologies en- 
courages spinning. We are accustomed to hearing competing 
interpretations from Democrats and Republicans, or conserva- 
tives and liberals, and statistics offer opportunities for spinning 
by these rivals. Thus, reports that a growing proportion of 
young Americans are overweight invite critiques from the left, 
targeting the food industry's campaigns to promote high-calorie 
products, and from the right, noting the obesity-enhancing 
effects of federal school lunch programs . 12 In most cases, the ar- 
guments chosen, the factors blamed for the problem, and the 
nature of the solutions proposed are predictable to anyone fa- 
miliar with the ideologies. 

The more figures available, the more opportunities for spin- 
ning. For example, the federal government collects extensive 
data regarding social problems such as drug use. Surveys of high 
school seniors, known as Monitoring the Future (MTF), provide 
one of the standard means for tracking drug use. Administered 
during most years, the MTF surveys generate statistics on sen- 
iors' self-reported use of different drugs over various periods of 
time. For example, we can learn that, in 2000, 21.6 percent of 
high school seniors reported smoking marijuana, and 50 percent 
reported drinking alcohol during the previous thirty days . 13 
What should we make of these numbers? Are things getting bet- 
ter or worse? It depends on which years and which drugs are 
used for comparison. For example, in 1990 — ten years earlier — 
14 percent of seniors reported smoking marijuana, so marijuana 
smoking was 50 percent higher in 2000; however, during the 
same period, drinking alcohol declined, from 57.1 percent to 50 
percent. Nor are the trends all that steady; almost every MTF re- 
port offers more than enough numbers to allow someone who 
picks figures carefully to argue, based on statistics, that teen drug 
use has either increased or decreased during a particular period. 

Such arguments are one form of cherry-picking (sometimes 
called data dredging) — that is, selecting statistics that support a 
particular thesis and drawing attention to those numbers, while 
ignoring other figures that might lead to a different conclusion. 
The amount of available data makes all the difference; the more 
numbers to choose among, the more certain one is to find some 
potentially useful "cherries," ripe for the picking. All manner of 
interested parties can adopt the tactic of cherry-picking. 
Political incumbents can point with pride to evidence of im- 
provements during their tenure in office, even as their chal- 

lengers argue that the facts show that things have deteriorated 
(and will likely get even worse unless the voters oust the rascals). 
Similarly, proponents of particular ideologies can select figures 
that seem to confirm their ideas. 

Without inspecting the original data, it can be hard to detect 
cherry-picking, although one suspicious sign is when advocates 
of some position offer very specific numbers in support of 
a broad argument. For example, someone might declare, 
"Between 1997 and 2000, the percentage of high school seniors 
who reported trying heroin during the previous thirty days rose 
by 40 percent!" While this is true (reported usage rose from 0.5 
to 0.7 percent, a 40 percent increase), the speaker ignores data 
from the same MTF reports showing that the seniors' reported 
use of most other drugs, including marijuana, cocaine, alcohol, 
and cigarettes, fell. But only the most careful listener might 
think to ask why the speaker chose to focus on one specific drug 
(particularly on one rarely used by high school students). 

Statistics, then, can be both the subjects of spinning, and — 
when carefully selected through cherry-picking — tools for 
spinners. Spinning may feature pretty good numbers, but be- 
cause these figures appear out of context, complexity and nuance 
have been stripped away. The statistics then can serve to pro- 
mote the viewpoint ofwhoever injects them into an ongoing de- 
bate. The point is not that some numbers are correct and others 
have been "spun"; rather, it is to caution us that every number 
presented in public debates may have been plucked from all the 
available figures because an advocate saw it as having rhetorical 
potential. Whenever numbers seem to offer especially powerful 
support to a particular position in a debate over a social issue, we 
need to be alert for signs of spinning or cherry-picking. 


Advocates who spin statistics recognize that numbers can have 
symbolic significance in debates over social issues. Answering 
even the simplest questions — How many? A lot or only a lit- 
tle? — can have powerful symbolic importance because differ- 
ent answers can seem to lend support to one side or another in 
social conflicts. Consider recent disputes over the size of two re- 
ligious populations — Muslims and Jews — in the United States. 
At first glance, the number of adherents to a particular religion 
might not seem like a topic that would generate intense interest; 
the number of Presbyterians, for instance, does not command 
much attention outside that denomination. But because the 
numbers of Jews and Muslims in this country may have impli- 
cations for how Americans think about the Israeli-Palestinian 
conflict, terrorism, and other foreign policy concerns, as well as 
about the future prospects for these religions in the United 
States, various groups have been bickering about both these 

The number of Muslims in the United States became a hot 
topic after the September it, 2001, terrorist attacks. Some wor- 
ried that the government or the public might blame all Muslims 
for the attacks, leading to a wave of anti-Muslim hate crimes or 
even repressive policies akin to the relocation of Japanese 
Americans during World War II. There also was a sense among 
some advocates that Muslims needed to be recognized as a sub- 
stantial religious minority within the United States, so that their 
interests and concerns might warrant more consideration. 

Recent estimates for the U.S. Muslim population range from 

fewer than two million to close to ten million. This may seem 
like a remarkable range, but remember (as noted in chapter i ) 
that the census — usually the most authoritative source for pop- 
ulation statistics — does not gather information on religion. 
Thus, it is necessary to find other ways to derive estimates. Some 
analysts have used national origin as a basis for calculations; 
they assume that people whose ancestors (or who themselves) 
came from largely Muslim countries are themselves Muslim. 
Others have tried to calculate the number of mosques in the 
United States, multiply that number by some average number 
of people affiliated with each mosque, and then add an estimate 
for Muslims unaffiliated with any mosque. Still other analysts 
derive their data from surveys that ask respondents to state their 
religion. Each of these methods has limitations. For example, 
religious affiliation among immigrants may not reflect the over- 
all pattern of religious affiliation among the population in their 
country of origin (just as immigrants to colonial Massachusetts 
were far more likely to be Puritans than the general English 
population). It is also difficult to identify all mosques, to judge 
estimates of the average number affiliated with a mosque, and 
to assess the estimates of unaffiliated Muslims. And, of course, 
because not everyone cooperates with surveys, survey results 
may undercount the populations they seek to measure.” 

The methods favored — and the critiques of rival methods — 
differ depending on one's position in the larger debate. Several 
major Muslim organizations, for example, with an understand- 
able interest in showing that their religion has many adherents, 
sponsored a study based on mosques. The study concluded that 
about two million people were associated with mosques and 
then, assuming that an even larger number of Muslims were not 

involved in mosques, argued that the overall Muslim popula- 
tion was between six and seven million. Some Jewish organiza- 
tions countered that this estimate was unreasonably high and 
pointed to survey results suggesting a total figure just below two 
million. Muslims, in turn, suggested that surveys undercount 
respondents who, for reasons of fear or language barriers, fail to 
report their religion to interviewers. One cannot help but sus- 
pect that the number of Jews (estimated at between five and 
six million) serves as an important benchmark in this debate: 
Muslim organizations favor estimates that place the number of 
Muslims as greater than the number of Jews, whereas Jewish 
organizations prefer figures that suggest that Jews outnumber 

Although recent efforts to estimate the size of the Muslim 
population attracted widespread interest, debate over the num- 
ber of Jews in the United States, while intense, has remained 
largely confined to the Jewish community. Here, the concern is 
not only that Jews are a small religious minority but also that 
their numbers may actually be declining, which has led some 
commentators to warn that American Jews may be "vanish- 
ing.'”' This concern, which has a long history, intensified after 
reports of recent social research. In both 1990 and 2000, major 
Jewish organizations sponsored the National Jewish Population 
Surveys (NJPS), large-scale research efforts designed to meas- 
ure the Jewish population. These surveys revealed that a slight 
majority of Jews were marrying non-Jews (raising the prospect 
that children from these marriages might not be raised as Jews) 
and that Jewish women bore an average of 1.8 children (that is, 
below the level needed to replace the Jewish population). 
Especially controversial was the news that the 2000 estimate for 

the Jewish population (5.2 million) was actually lower than the 
1990 figure (5.5 million). Critics charged that the NJPS had 
badly undercounted the Jewish population, that the correct total 
was closer to 6.7 million. 

This debate hinged on disagreements about how best to 
define who is Jewish. Is it a matter of religious practice? Or are 
Jews those who think of themselves as Jewish? The NJPS 
counted both categories. But what about people who say they 
were once but are no longer Jewish, or who live in households 
with others who report being Jewish? People in these cate- 
gories, excluded by the NJPS, were counted in the critics' esti- 
mates. In addition, the critics argued that fear of anti-Semitism 
would have led some NJPS respondents to deny being Jewish. 
As always, a narrow definition will produce a smaller estimate 
than a broad definition. 

At one level, debates over estimates of the numbers of Mus- 
lims and Jews in the United States can be seen as questions of 
method and definition. If we try to count Muslims by estimat- 
ing how many people are affiliated with mosques and then add 
an estimate for the unaffiliated, we get one (high) number; if we 
use survey research, we get a second (lower) figure. Similarly, 
defining Jewish identity narrowly produces a lower estimate 
than adopting a broader definition. On a technical level, social 
scientists can debate the advantages and limitations of the 
different methods and definitions. (The consensus would prob- 
ably favor using surveys to estimate the Muslim population and 
the narrower NJPS definition to identify Jews, but not everyone 
would agree.) But, of course, these are not merely technical 
questions. These statistics have symbolic importance — to argue 
that a group is large or growing may suggest that its interests 

should be more important than those of a group that is smaller 
or shrinking. And commitment to such political messages can 
lead to impassioned defenses of numbers that might not receive 
strong support on purely technical grounds. 


The 1996 federal welfare reform law (the Personal Responsi- 
bility and Work Opportunity Reconciliation Act, or PRWORA) 
was the product of decades of bitter debate. Most often, critics 
of welfare complained that the system fostered long-term de- 
pendency, that some recipients not only remained on welfare 
but also raised children who, in turn, would themselves spend 
their adult lives on welfare, in a troubling intergenerational 
cycle. The central argument was that welfare discouraged self- 
reliance and personal responsibility and that it had become self- 
perpetuating. In contrast, defenders of welfare insisted that it 
was necessary, that it was a vital safety net providing minimal 
protection for individuals who had too few resources to provide 
for themselves in a society that otherwise offered limited oppor- 
tunities. Rather than seeing welfare recipients as individuals 
who failed to exercise responsibility, they blamed a social system 
that featured too few jobs and too much discrimination. 

Both welfare’s critics and its defenders understood that wel- 
fare was linked to other social problems: recipients were, by 
definition, poor; in addition, they tended to have less education 
and more serious health problems than those not receiving wel- 
fare. Many women on welfare bore their children out of wed- 
lock, and the fathers of their children often lacked jobs. To the 
critics, welfare discouraged work and marriage; to the defend- 

ers, the absence of decent employment opportunities created the 
social circumstances that forced people to turn to and stay on 
welfare. Defenders called for more benefits to improve recipi- 
ents' standard of living, as well as job training and other pro- 
grams to improve their prospects, while critics charged that 
raising benefits and expanding programs only fostered depen- 
dency. The debate stretched over decades. 

PRWORA — the product of a Republican-controlled Con- 
gress and signed by President Bill Clinton — was designed to 
"end welfare as we know it." In particular, the new law replaced 
the old Aid to Families with Dependent Children (AFDC) enti- 
tlement with Temporary Assistance for Needy Families (TANF) 
block grants to the states, giving the states considerable discretion 
in designing arrangements to help the poor. In addition, the new 
law established a lifetime limit for most recipients of no more 
than sixty months of cash assistance from federal funds, and it 
required that recipients work after receiving two years of cash 

These were billed as significant, dramatic changes. The new 
law's supporters (mostly conservatives) envisioned a rosy future 
in which the formerly dependent would learn personal respon- 
sibility, take charge of their lives, and work their way up from 
poverty. Its critics (mostly liberals) warned of an impending so- 
cial catastrophe in which declining support would force mil- 
lions of people into poverty and increase the ranks of the home- 
less. Supporters promoted the new law as providing encourage- 
ment; critics insisted that it would be harsh and punitive. 

The fifth anniversary of the welfare reform legislation — a 
significant date because it marked the passing of sixty months 
(PRWORA s lifetime limit for cash assistance) — provided an 

occasion to assess the law's impact. 16 Supporters presented nu- 
merous statistics as evidence that welfare reform had been a 
success: the number of households receiving assistance was 
down from AFDC’s 1994 peak of 5.1 million to about 2 million; 
the proportion of single mothers who were working had risen; 
and the proportion of births to unmarried mothers — which 
had risen rapidly during the years just preceding welfare re- 
form — had barely increased. Even PRWORAs most vigorous 
critics had to concede that, at least so far, the predicted catastro- 
phe had not occurred. 

What had happened? Analysts concluded that the overall 
statistics were affected by several developments. The first was 
good timing: the early years of welfare reform occurred during 
an economic boom, when unemployment was low and jobs 
were relatively plentiful. Second, and probably more important, 
although this effort was less publicized than welfare reform, the 
federal government had instituted or expanded policies that 
were designed to assist low-income workers, such as the Earned 
Income Tax Credit (EITC), increased aid for child care, and ex- 
panded access to Medicaid. Since much of the public hostility to 
welfare centered on the recipients' failure to work, these pro- 
grams to assist people who were working had much broader 
support. The increased benefits provided by these programs 
meant that at least some low-income workers who in the past 
might have been forced to go on welfare to qualify for needed 
support (say, to deal with a child's medical bills) now could re- 
main employed. 

Much was made of the number of recipients who left the wel- 
fare rolls, but this was in some ways a selective reading of the ev- 
idence. It had always been the case that most recipients received 

only short-term aid; people had been going off welfare during 
every year of the program's history. It was difficult to establish 
whether there had been a marked increase in the numbers of 
people who left the welfare rolls — post-PRWORA record- 
keepers carefully counted their numbers, but the records of ear- 
lier years were less complete. The real difference may have been 
a third factor: PRWORA, which allowed states to set standards 
for eligibility, made it harder to qualify for welfare benefits. 
Instead of being added to the welfare rolls in the first place, 
would-be recipients might be urged to apply for jobs or required 
to fulfill other requirements that had the effect of encouraging 
them to consider options other than going on welfare. 

Different commentators weighed these factors differently, in 
fairly predictable ways. Welfare reform's critics tended to em- 
phasize the importance of the healthy economy for the pro- 
gram's apparent success (and they watched with foreboding as 
the economic boom ended). They also noted the importance of 
the various programs to support the working poor (which were 
generally more popular with liberals than with conservatives). In 
contrast, PRWORA s supporters argued that the program's suc- 
cess revealed that welfare had been unnecessary in many cases, 
and they called for new reforms to restrict benefits further. 

Still other critics suggested that the overall assessment of suc- 
cess overlooked evidence indicating that welfare reform had had 
harmful consequences for some of the poor. In 2003, for exam- 
ple, the Children's Defense Fund (CDF) noted that, although 
the proportion of black children living in families officially 
defined as poor had dropped markedly since 1995, the number 
of black children being raised in "extreme poverty" (which the 
CDF defined as households with incomes no more than half the 

federal poverty line) increased sharply between 2000 and 2001 , 17 
The report generated sympathetic editorials about the plight of 
"the poorest of the poor" in some newspapers as well as accusa- 
tions of cherry-picking. (One conservative charged that the CDF 
had "searched with a laser for something negative to say, because 
the poverty picture in America since the 1996 welfare reform is 
unambiguously positive.") But, of course, we needn't presume 
that a policy change such as welfare reform will have the same 
effect on every person. Complicated problems are not likely to 
have simple solutions, any more than they are likely to have sim- 
ple causes. 

Social policies such as PRWORA are relatively blunt instru- 
ments designed to address problems that usually involve compli- 
cated tangles of causes and consequences. When people have dis- 
agreed over whether introducing a policy would be desirable or 
ill advised, they are unlikely to agree about its effects. Again, the 
notion of trade-offs is relevant: every policy is likely to have 
benefits and costs, and its proponents can be expected to praise 
the benefits, even as their opponents decry the costs. This com- 
plexity should make us wary of simplistic policy assessments that 
pronounce success or failure, as evidenced by one or two statis- 
tics. Complexity cannot be summarized in a couple of magical, 
cherry-picked numbers, and we should question claims that 
make everything seem simple and straightforward. 


Americans have a widespread, naive faith in the power of num- 
bers to resolve debates, to provide facts that can overpower op- 
position. This faith rests on some dubious assumptions. The 

first is a belief that numbers are by nature factual, that they con- 
stitute incontrovertible evidence. This ignores an even more 
basic truth — that all numbers are products of human efforts. 
We cannot escape the fact that statistics are social constructions. 

Recognizing this means that we can't treat numbers as 
straightforward bits of truth; rather, we must be critical, asking 
who counted what, and how, and why. But it does not mean 
that we can't trust any statistics, that we should treat them all as 
equally worthless. There are better and worse ways of counting, 
and we can have more confidence in some numbers than in oth- 
ers. All science is not junk science; with a little effort — and the 
patience to wait for more information — we can distinguish be- 
tween the two. 

A second weak assumption is that our side's numbers are bet- 
ter than the other side's numbers, simply because they're ours. 
Our positions, biases, political ideologies, and perspectives shape 
how we approach evidence, including statistics. We have a nat- 
ural tendency to welcome numbers that reaffirm what we be- 
lieve to be true. Precisely because these figures are consistent 
with our view of the world, we tend to downplay — if not be 
oblivious to — their weaknesses. We give them a sympathetic 
reading, a free ride. In contrast, our critical faculties swing into 
operation when we confront numbers that challenge our beliefs. 
Now we ask the hard questions: What could have led to num- 
bers that are so obviously wrong? Was it peculiar definitions? 
Faulty methods? Bad samples? Inappropriate analysis? There 
is nothing like a discomforting statistic to help most of us un- 
cover critical abilities we might not have realized we had. 

This chapter also suggests a third consideration: social life is 
complicated. While discussions of some social problems are 

one-sided (child molestation and serial murder have few de- 
fenders), other issues lead to debates among people with differ- 
ent assumptions, beliefs, attitudes, and values. Debaters com- 
monly unveil statistics that support their positions, and of 
course they find their own numbers convincing, even as they ex- 
press deep reservations about their opponents' figures. Whether 
we are actively engaged in these debates or somewhere on the 
sidelines, it may help to consider the possibility of complexity. 
The choices that people make in counting are necessary and are 
undoubtedly shaped by many factors — some methods of count- 
ing are cheaper than others, some are a better fit for the people 
doing the counting, some may seem more likely to lead to the 
results they hope to find. We should expect that the choices peo- 
ple make shape their results. 

We shouldn't presume that most social issues are simple to 
understand. If they were all that simple, we wouldn't have all 
that disagreement. When we encounter disagreements about 
the validity of a number, or when we hear people promoting 
rival numbers, we ought to consider the possibility that there 
may be an underlying complexity — that instead of trying to de- 
cide which side owns the truth, we might be better off trying to 
reconcile the competing claims, to understand how and why 
people have different visions of what's true. We may, of course, 
decide in favor of one position, but then we may also come to 
understand that it's a little more complicated than that. 


B ad statistics aren't rare. You can probably spot at least 
one dubious number in this morning's newspaper. 
Recognizing bad statistics is not all that difficult; it 
takes clear thinking more than it requires any 
advanced mathematical knowledge. And most people will agree 
that we ought to stamp out bad statistics. 

Still, bad numbers flourish. Why? Shouldn't we be able to 
teach "statistical literacy" — basic skills for critically interpret- 
ing the sorts of statistics we encounter in everyday life? Why 
can't statistical literacy be part of the standard high school or 
college curriculum? Shouldn't we be able to, in effect, immu- 
nize young people so that they will be able to think critically 
about the numbers they encounter and resist bad statistics? 


Every year, thousands of high school seniors enroll in Advanced 
Placement statistics classes. (At the end of the year, these stu- 
dents can take the national AP statistics exam, and, if they score 
well enough, many colleges will give them credit for having 
completed a basic statistics course.) Many thousands more stu- 
dents will take at least one statistics course in college. We might 
expect that statistical literacy would be an important part of 
these courses. 

We would be wrong. Statistics textbooks, as well as the AP 
exam, all but ignore the sorts of issues raised in this book. 
Rather, statistics instruction, in both high school and college, fo- 
cuses on what I call matters of calculation — on the theory and 
logic behind particular statistical measures, on the methods of 
actually computing those measures, and on the interpretation of 
the results. Introductory statistics textbooks feature chapters on 
probability theory, on tests of significance, correlation, regres- 
sion, and so on. That is, these textbooks assume that the students 
who read them might want to use statistics to interpret data de- 
rived from some sort of scientific research. There is nothing 
wrong with this; those students who do become researchers will 
indeed need to know how to calculate those statistics. 

However, such textbooks and courses say next to nothing 
about how to interpret the simple statistics — the graphs and 
numbers — the students might encounter in the morning news- 
paper. Why? If everyone agrees that statistical literacy is an 
important skill, why isn't it an important part of statistics 


1 7 1 

Statistics is usually understood as a branch of mathematics, 
hence the focus on calculation. I am sure that most high schools 
consider the AP statistics class to be a math class and assign a 
math teacher to teach it. The goal of the course is to make stu- 
dents proficient statistical calculators; the classes are not de- 
signed to make them statistically literate. To no one's surprise, 
math teachers believe that their job is to teach math, to teach 
students how to calculate correctly so that the students can score 
well on tests of calculation, such as the AP statistics exam. 

Similarly, the statistics courses taught in college — even the 
basic, introductory courses — devote almost all their attention to 
matters of calculation. The spread of computers and easily mas- 
tered statistical software packages has encouraged the use of 
highly sophisticated statistics. Before 1970 or so, a person with- 
out advanced training in statistics who picked up an issue of a 
leading social science journal, such as the American Sociological 
Review, could probably understand the data presented in many 
of the articles. This is no longer true. Today's A SR articles fea- 
ture ordinary least- squares regression, log-linear regression, 
and other complex, multivariate statistical techniques that prob- 
ably cannot be understood by anyone who has not taken at least 
two semesters of statistics in college. Naturally, college instruc- 
tors believe that their job is to teach students to master these ad- 
vanced techniques. 

Could it be that the kinds of issues I've raised in this book 
strike most statistics instructors (and textbook authors) as too 
simple to warrant comment? Perhaps. But more than that, the 
topics we've covered aren't matters of calculation. We have been 
less concerned with mathematical processes (calculations) than 
with a social process. Our focus has been on who counts — who 

1 72 


produces numbers, why they produce them, which audiences 
consume them, and how those numbers are understood and put 
to use. That is, we have tried to understand the social construc- 
tion of numbers more than their calculation. 

But statistics classes largely ignore the ways statistics are used 
as evidence for understanding social issues as well as the ways 
people count. If the social process by which statistics are brought 
into being is mentioned, it is probably in relation to the idea of 
bias — instructors may warn students that "biased" people can 
devise distorted statistics. But beyond blaming bias — which is 
treated as a sort of contamination originating outside the math- 
ematically pure realm of calculation — statistics classes rarely 
explore what this distortion might involve. 

In short, even if everyone agrees that it would be desirable 
for students to improve their ability to think critically about 
the sorts of statistics found in news coverage, statistics teachers 
aren't likely to feel that this is their job. 


Contemporary educators are beset by competing demands. On 
the one hand, as new social issues come to public attention, there 
are often calls to add material to the schools' curriculum; sex ed- 
ucation and drug education are obvious examples, among many 
others. A school district may win a grant for an anti-bullying 
program. There may be campaigns at the state or school district 
level to make students aware of various sorts of discrimination. 
The list goes on and on, and it changes with each passing year. 
Some of these new special topics become enduring elements in 

1 73 


the schools' curriculum, but others turn out to be short-lived en- 
thusiasms, educational fads. 

On the other hand, many grumble that schools are neglect- 
ing the basics, the Three Rs. The school accountability move- 
ment, at least in part, demands that schools return to emphasiz- 
ing instruction in basic skills. Schools and teachers, then, find 
themselves trapped between calls to spend more time teaching 
basic skills and pressure to add instruction about whatever new 
special topics currently occupy the public's attention. The school 
day contains only a limited number of minutes, and all sorts of 
people want more minutes to be devoted to whatever topics they 
deem important. 

So a first question might be whether statistical literacy ought 
to be considered an additional special topic or a basic skill. If it 
is promoted as a special topic — like AIDS education and bully- 
ing prevention — its long-term prospects won't be bright. This 
year's addition to the curriculum easily becomes a candidate for 
elimination when next year arrives with its calls to teach still 
other new topics. 

Well, what if we call statistical literacy a basic skill? Cer- 
tainly a plausible argument exists for considering it in these 
terms. After all, we are talking about teaching people to be 
more critical, to be more thoughtful about what they read in the 
newspaper or watch in a news broadcast, to ask questions about 
claims from scientists, politicians, or activists. Being better able 
to assess such claims is certainly valuable; we might even argue 
that it is fundamental to being an informed citizen. Why not 
consider statistical literacy a basic skill? 

But this raises another question: what sort of basic skill is it? 
The answer matters because both high schools and colleges par- 



cel out responsibility for instruction to departments organized to 
teach topics. A typical high school has separate departments for 
science, social sciences, mathematics, English, and so on; most 
colleges subdivide many of these broad categories, for example, 
assigning the responsibility for teaching to separate departments 
for biology, chemistry, and so on. In general, the larger the edu- 
cational institution, the more departments it recognizes. 

Departments are natural competitors. While everyone may 
acknowledge the value of a well-rounded education, each de- 
partment tends to assume that it plays an especially important 
role. And because money is always short, departments compete 
for available funds to hire faculty and purchase equipment. It is 
the rare department that doesn't want to expand; in particular, 
many departments would like to offer more advanced training, 
such as AP courses in high schools or graduate programs in 

This competition means that teaching basic skills often is de- 
valued. For example, almost all of the thousands of first-year 
students admitted to large universities each year are required to 
take an English composition class. Those classes need to be 
small, because the students must write a lot of papers, and those 
papers need to be graded quickly and carefully. At most univer- 
sities, the job of teaching those composition classes falls on grad- 
uate students or part-time instructors, not on English profes- 
sors. In part, it is much cheaper to teach composition this way; 
in part, English professors prefer to teach advanced courses 
to English majors (because both the subject matter and the stu- 
dents are more interesting). The point is that teaching this basic 
skill is not considered particularly rewarding. (Some univer- 
sities' English departments have spun off separate departments 

1 75 


of composition, writing centers, or other programs to handle 
this unpleasant chore.) 

The example of English composition can help us appreciate 
the problems of teaching statistical literacy. College instructors 
are well aware that substantial proportions of students have 
trouble reading — let alone thinking critically about — basic 
graphs or tables. This is a very important skill because graphs 
and tables are certain to appear in much of the reading a student 
will need to do in the course of college. And yet, no one wants 
to teach this skill, or at least to spend much time doing so. Many 
have the sense that students should already be proficient in these 
skills when they get to college (even though it is clear that many 
are not). To many others, it seems too simple, too basic — a 
waste of time for professors who would prefer to teach the more 
advanced topics in their disciplines. 

In addition, the spread of personal computers and sophisti- 
cated software helps sustain the illusion that students already 
understand this stuff. Anyone who visits a junior high school 
science fair will see all manner of eye-catching, computer- 
generated graphs. As long as no one bothers to ask whether 
these graphs are clear and useful (they often are neither), it is 
easy to be impressed by what the students have produced. Simi- 
larly, students learn that they can find answers to pretty much 
any question by searching the Internet. They may not locate 
particularly good answers, but they find answers all the same. 
The experience that many students already have in using high- 
tech methods (albeit to produce low-quality results) helps to jus- 
tify claims that we don't need to teach basic skills, that we can 
move on to teaching more interesting, advanced material. 

Thus, statistics and mathematics instructors are unlikely to 



have any more interest in teaching statistical literacy than 
English professors have in teaching first-year composition. Nor 
are other departments eager to teach this material. I teach soci- 
ology courses, but I know that most sociology professors tend to 
dismiss statistical literacy as "not really sociology"; faculty in 
psychology and other disciplines probably have the same reac- 
tion. Statistical literacy falls between the stools on which aca- 
demic departments perch. 

There is precedent to support my pessimism. During the late 
1980s and early 1990s, "critical thinking" became a buzzword 
on college campuses. This should have been the perfect slogan 
around which to rally support for educational reform. Virtually 
all professors consider themselves critical thinkers, and most 
would agree that students must learn to think more critically — 
another highly desirable basic skill. But because all those pro- 
fessors believed that they already were teaching their students to 
think critically (even though they simultaneously complained 
that many students were poor critical thinkers), and because no 
department wanted to take on the responsibility for teaching 
the topic across the campus, interest in improving critical think- 
ing peaked, and the strength of the idea as an educational slo- 
gan has begun to fade. 

What happened to critical thinking? Why didn't that good 
idea become an enduring part of education in all schools? The 
lack of a departmental "owner," a department that would 
house, protect, and nurture critical thinking, meant that teach- 
ing the skill remained everyone's responsibility — and therefore 
no one's. 

This example suggests that a specific department needs to 
take responsibility for teaching statistical literacy. As we have 

1 77 


already established, this is not likely to be a mathematics or sta- 
tistics program, however logical that might seem at first glance. 
The social sciences might offer an alternative home. After all, 
issues of statistical literacy often emerge around discussions of 
social issues. But again, sociology professors are likely to dismiss 
statistical literacy as not being "real sociology" (and other de- 
partments may react the same way). 

Departmental organization offers considerable advantages 
for educational institutions, but it also carries costs. It is difficult 
to teach subjects that do not fit neatly within what a department 
considers its proper instructional domain. This helps to explain 
why many graduates of high schools and colleges remain un- 
comfortable when confronted with even basic statistics — and 
why this situation will not change easily. The lessons involved 
in teaching statistical literacy are not so terribly difficult; rather, 
the difficulty hes in finding someone willing to teach them. 


Despite these obstacles, a small educational movement advocat- 
ing statistical literacy has emerged. Professor Milo Schield, 
director of the W.M. Kleck Foundation Statistical Literacy 
Project at Augsburg College in Minneapolis, is the movement's 
leading voice. Schield operates the Statistical Literacy Web site 
(; for those interested in statistical literacy as 
an educational movement, the site includes a section on teach- 
ing. Although this is a promising development, the campaign to 
promote formal instruction in statistical literacy is in its early 

But perhaps statistical literacy doesn't have to be taught in 

1 78 


classrooms. Recently, there seem to be increasing calls to pro- 
mote statistical literacy outside the educational establishment. 
Consider, for example, these resources: 

The Statistical Assessment Service ( has been 
criticizing the media's handling of statistics since 1995- SAS pub- 
lished newsletters until 2002, when it converted to distributing 
its reports on its Web site. A book based on SAS analyses is both 
readable and available in paperback; see David Murray, Joel 
Schwartz, and S. Robert Lichter, It Ain't Necessarily So: How 
Media Make and Unmake the Scientific Picture of Reality (2001). 

Various Web sites from around the world feature discussions 
of bad statistics. Some of these contain mostly original material; 
others are little more than links to specific discussions around the 
Web. Numberwatch ( is a British site; 
its operator, John Brignell, is the author of Sorry, Wrong Number! 
The Abuse of Measurement (2000). The Social Issues Research 
Centre ( is another British site presenting analy- 
ses of issues that often involve critiques of statistics. Ptnombre 
( is a French site, which also contains some 
materials in English. The Canadian Statistical Assessment Ser- 
vice ( resembles its U.S. counterpart, while an- 
other Canadian site, (, is 
basically a catalog of links. Numeracy in the News, an Australian 
site, is aimed at educators and students; it features sample articles, 
graphs, and so on, each accompanied by study questions and com- 
mentary ( 
.htm). Many of these organizations also offer links to more spe- 
cialized sites, including official statistics (many government agen- 
cies now provide sites where one can access their statistical 


reports) and sites devoted to particular social issues or types of 
data — for example, Quackwatch ( on 
medical claims, ( on 
media coverage of scientific news, and the Center for Media and 
Democracy ( for critiques of industry and gov- 
ernment public relations campaigns. As might be expected, such 
sites vary in their concerns and underlying ideologies, and their 
critiques should be examined critically rather than simply being 

. It's often fun to explore bad statistics, but for sheer enter- 
tainment, it is hard to beat Cecil Adams's column, "The Straight 
Dope," which appears in alternative weekly newspapers. Its 
motto is "Fighting Ignorance Since 1973 (It's Taking Longer 
Than We Thought)." Each week, Adams addresses one or more 
questions — often on topics that good taste leads other media to 
ignore; some, although by no means all, involve sorting out sta- 
tistical claims. The Web site ( offers an 
index for and access to all the columns. If you're interested in 
exotic topics, this is a wonderful resource. 

Other media commentators also promote statistical literacy. 
The mathematician John Allen Paulos, author of Innumeracy: 
Mathematical Illiteracy and Its Consequences (200 1) and other 
books for general readers, has a Web site (http://euclid.math -paulos/) that links to his various works, including 
his columns for The British Broadcasting Corpora- 
tion has several mathematically themed radio programs, includ- 
ing "More or Less," which features frequent commentaries on 
statistical issues. Broadcasts are archived at 


. The American Statistical Association publishes Chance, a 
quarterly magazine devoted to interesting uses of statistics. 
Some of the articles require considerable background in statis- 
tics, but others are more accessible. As an introduction to what 
professional statisticians do, it is a valuable resource. 

. Many books on statistical topics are available, ranging from 
textbooks that teach students how to calculate different statistics 
to volumes — such as this one — that offer critiques of how sta- 
tistics are used and misused in contemporary society. (Several of 
these books are listed in the notes to earlier chapters of this 

These various sources form a chorus of voices promoting the 
cause of statistical literacy. Of course, disagreements arise with- 
in the movement. Some advocates have ideological agendas: 
conservatives concentrate on exposing liberals' misuse of statis- 
tics, while liberals attack dubious numbers promoted by conser- 
vatives. Some critics seem to blame "the media" for irresponsi- 
bly publicizing bad statistics, but journalists — not unreason- 
ably — respond that they often have no good way to assess the 
numbers their sources offer. Some statisticians advocate better 
mathematical training to improve our understanding of calcu- 
lation, while social scientists (such as myself) argue that it is im- 
portant to locate numbers within the social context that creates 
and disseminates them. 

In short, it may be true that "everyone" agrees that improving 
statistical literacy is desirable, but it isn't clear that they can agree 
on what statistical literacy means, what improving it might in- 
volve, or what the consequences of this improvement might be. 


1 8 1 


Even if no one opposes statistical literacy, serious obstacles re- 
main. There is disagreement about which skills need to be 
taught, and, at least so far, no group has offered to take respon- 
sibility for doing the necessary teaching. Plenty of information 
is out there — any interested individual can learn ways to think 
more critically about statistics — but the statistical literacy 
movement has yet to convince most educators that they need to 
change what the educational system is doing. 

Many of us kid ourselves that bad statistics come from people 
with whom we disagree, and we fantasize that improving statis- 
tical literacy will inevitably swell the ranks of people who agree 
with us, that all critical thinkers will recognize the flaws in our 
opponents' arguments, while finding our claims convincing. 

I wouldn't count on things working out that way. Statistical 
literacy is a tool, and, like most tools, it can be used for many 
purposes. If more people think more critically about statistics, 
they are likely to use that skill to criticize our numbers as well 
as those of our opponents. When everyone's numbers come 
under scrutiny, we are all held to higher standards. 

But that's not bad. As things stand, we constantly find our- 
selves exposed to lots of statistics. Some of those numbers are 
pretty good, but many aren't. As a result, we worry about things 
that probably aren't worth the trouble, even as we ignore things 
that ought to warrant our attention. Improving statistical liter- 
acy — if we can manage it — could help us tell the difference 
and, in a small way, make us wiser. 





1. Some readers may recall that I made a similar statement in 
Damned Lies and Statistics: Untangling Numbers from the Media, Politi- 
cians, and Activists (Berkeley: University of California Press, 200 1), p. 27. 
This book presents different examples, organized in a different way, 
and I have tried to minimize the overlap, but the underlying approach 
is the same. Youneedn't have read the first book tounderstand this one, 
however; the two are intended to complement each other. 

2. Darrell Huff, How to Lie with Statistics (New York: Norton, 
1954); Gerald E. Jones, How to Lie with Charts (San Francisco: Sybex, 
1995); Mark Monin.oniex ? How to Lie with Maps (Chicago: University of 
Chicago Press, 1991); Robert Hooke, How to Tell the Liars from the Sta- 
tisticians (New York: Marcel Dekker, 1983); Richard P. Runyon, How 
Numbers Lie (Lexington, Mass.: Lewis, 1981); Best, Damned Lies and 
Statistics; also: Tukufu Zuberi, Thicker than Blood: How Racial Statistics 
Lie (Minneapolis: University of Minnesota Press, 2001). Other books 
have chapters on the theme: see, for example, "Statistics and Damned 
Lies," in A. K. Dewdney, 200% of Nothing (New York: Wiley, 1993), 

pp. 23— 42. In addition, a recent three-hundred-page book designed to 
help statistics instructors reach their students includes a seventeen-page 
chapter entitled "Lying with Statistics," which addresses issues of bias; 
see Andrew Gelman and Deborah Nolan, Teaching Statistics: A Bag of 
Trices (New York: Oxford University Press, 2002), pp. 147-163. 


1. Rather is quoted in Lynnell Hancock, "The School Shootings: 
Why Context Counts," Columbia Journalism Review 40 (May 2001): 76. 
Much has been published about these episodes, and particularly on the 
Columbine shootings. See, for example, Daniel M. Filler, "Random 
Violence and the Transformation of the Juvenile Justice Debate," Vir- 
ginia Law Review 85 (2000): 1095— 1125; and Gary Kleck, "There Are 
No Lessons to Be Learned from Littleton," Criminal Justice Ethics 18 
(Winter 1999): 2, 61— 63. 

2. Relevant compilations of evidence regarding school violence 
include Margaret Small and Kellie Dressier Tetrick, "School Violence: 
An Overview," Juvenile Justice (U.S. Office of Juvenile Justice and 
Delinquency Prevention) 8 (June2ooi): 3—12; Phillip Kaufman et ah, 
Indicators of School Crime and Safety: 2001 (Washington, D.C.: U.S. 
Departments of Education and Justice, 2001); Kim Brooks, Vincent 
Schiraldi, and Jason Ziedenberg, School House Hype: Two Years later 
(Washington, D.C.: Justice Policy Institute, 1999); National School 
Safety Center, School Associated Violent Deaths, 2001, available at www; and Mark Anderson et al., "School- Associated Violent 
Deaths in the United States, 1994-1999 f Journal of the American Med- 
ical Association 286 (December 5, 2001): 2695-2702. For information 
about journalists' standards for assessing crime statistics, see Kurt Sil- 
ver, Understanding Crime Statistics: A Reporter's Guide (n.p.: Investi- 
gative Reporters and Editors Inc., 2000). 

3. Chip Heath, Chris Bell, and Emily Sternberg, "Emotional Selec- 
tion in Memes: The Case of Urban Legends f Journal of Personality and 
Social Psychology 81 (2001): 1028—1041. 

4* Cynthia J. Bogard, Seasons Such as These (Hawthorne, N.Y.: 
Aldine de Gruyter, 2003). 

5. Compare Kiron K. Skinner, Annelise Anderson, and Martin 
Anderson, Reagan f in His Own Hand (New York: Free Press, 200 1), pp. 
241, 459; and Tip O'Neill, Man cf the House (New York: Random 
House, 1987), pp. 347-348. 

6. In response, welfare critics offered their own numbers suggest- 
ing that abuses were widespread. One analysis suggested: "A major 
source of the variations among estimates [for welfare fraud] . . . has 
been the estimators' use of very different definitions of improprieties" 
(JohnA. Gardiner and Theodore R. Lyman, The Fraud Control Game 
[Bloomington: Indiana University Press, 1984], p. 2). 

7. On the 1980s, see Joel Best, Threatened Children (Chicago: Uni- 
versity of Chicago Press, 1990). On 2002, see Donna Leinwand, "Kid- 
napping Problem 'Impossible' to Quantify," USA Today, August 15, 
2002, p. 3A. 

8. For an introduction to this approach, see Anthony E. Boardman 
et ah, Cost-Benefit Analysis: Concepts and Practice (Upper Saddle River, 
N.J.: Prentice Hall, 1996). 

9. For a critique of one oft-repeated tale of amoral cost-benefit ana- 
lysts, see Matthew T. Lee and M. David Ermann, "Pinto 'Madness' as 
a Flawed Landmark Narrative," Social Problems 46 (1999): 30—47. 

10. Viviana A. Zelizer, Pricing the Priceless Child (New York: Basic 
Books, 1985). 

11. William Petersen, Ethnicity Counts (New Brunswick, N.J.: 
Transaction, 1997), pp. 78— 81. 

12. C. Kirk Hadaway, Penny Long Marler, and Mark Chaves, 
"What the Polls Don't Show: A Closer Look at U.S. Church Atten- 
dance," American Sociological Review 58 (1993): 741—752. 

13. On the 2000 census, see Margo Anderson and Stephen E. Fien- 
berg, "Census 2000 and the Politics of Census Taking," Society 39 
(November 2001): 17—25; and Eric Schmitt, "For 7 Million People in 
Census, One Race Category Isn't Enough,” New Yor ^ Times f March 1 3, 
2001, p. A 1 . For historical background, see Petersen, Ethnicity Counts. 

14* For the basic arguments about how to measureunemployment, 
see David Leonhardt, "Breadline? What Breadline?” New YoH { Timcc, 
June 24, 2001, sec. 4, p. 5; and Leonhardt, “Help Wanted: Out of a Job 
and No Longer Looking," New York Times, September 29, 2002, sec. 4, 
p. 1. 

15. Fox Butterfield, "When Police Shoot, Who's Counting?" New 
Yor^Timcc, April 29, 2001, sec. 4, p. 5. For a more general critique of 
how politics can shape what officials choosenot to count, see A. P. Tmt, 
"The Politics of Official Statistics," Government and Opposition 30 
( : 995) : 254-266. 

16 . For a case study of how political considerations can shape the 
ways an official agency (the Canadian Centre for Justice Statistics) col- 
lects, analyzes, and presents data, see Kevin D. Haggerty, Making 
Crime Count (Toronto: University of Toronto Press, 2001). 

17. American Medical Association, Office of Alcohol and Other 
Drug Abuse, "College Binge-Drinking Prevention Program Calls on 
Princeton Review to Stop Publishing 'Party Schools List,"' pressrelease, 
August 7 , 2002. 

18 . Bureau of the Census, Historical Statistics cf the United States 
(Washington, D.C., 1975), p* 58 . 

19. On some of the ways progress serves to make us more conscious 
of social problems, see Joel Best, "Social Progress and Social Problems," 
Sociological Quarterly 42 (2001): 1 — 12. 

20. The spread of the falling-coconuts statistic was aided by a 
British travel insurer; see Beverly Beckham, "Travelers Should Watch 
Out for Coconuts: The Killer Fruit," Boston Herald, April 7 , 2002, 
p. 30. A discussion of the available data appeared on Cecil Adams's 
splendid Web site "The Straight Dope" on July 19, 2002; see www 9. 

21. Lawrence K. Altman, "Stop Those Presses! Blonds, It Seems, 
Will Survive After All," New York Timcc, October 2, 2002, p. A5. 

22. Robert Schoen and Robin M. Weinick, "The Slowing Metabo- 
lism of Marriage: Figures from 1988 U.S. Marital Status Life Tables," 
Demography 30 (1993): 737—746. The earliest statement of this erro- 

neous statistic I have found appears in Judy Klemesrud, '"If Your Face 
Isn't Young': Women Confront Problems of Aging," New Yod { Timcc, 
October io, 1980, p. A24. 

23. Edward M. Eveld, "Awash in Water: Eight Glasses a Day? It's 
Probably More Than We Need, Scientists Say," Seattle Timcc, July 13, 
2003, p. L6. 

24. For a thorough discussion of this example, see Jonathan Marks, 
What It Means to Be 98% Chimpanzee: Apes , People, and Their Genes 
(Berkeley: University of California Press, 2002). 


1. Cal Thomas, "Democrats Are Losing Center," Wilmington (Del) 
News Journal \ August 3, 200 3, p. An. 

2. For examples of claims made by proponents and opponents of 
the tax cut, see David E. Rosenbaum, "Washington Memo: The Presi- 
dent's Tax Cut and Its Unspoken Numbers," New Yorl^ Times, Febru- 
ary 25, 2003, p. A 25. 

3. For an introduction to the concept, see John Allen Paulos, Innu- 
meracy: Mathematical Illiteracy and Its Consequences (New York: Ran- 
dom House, 1988). 

4. Issues of definition and measurement are discussed in Best, 
Damned Lies and Statistics t pp. 39— 52. 

5. After inventing this example, I discovered that studies of the 
relationship between sinistrality (left-handedness) and delinquency do 
exist. Some report evidence that the two are related (although not as 
powerfully as my imaginary data suggest), but there seems to be no 
agreement about why this might be true. See, for instance, William C. 
Grace, "Strength of Handedness as an Indicant of Delinquents' Behav- 
ior J Journal of Clinical Psychology 43 (1987):igi — 155. 

6. It is usually legitimate, however, to compare two or more 
changes, expressed as percentages, that occurred during the same 
period. That is, ifwemeasure the rates for two different crimes in 1980 
and again in 1990, and our data show that crime X rose 25 percent, 

while crime Y rose 50 percent, we can conclude that crime Y rose faster 
than crime X. 

7. For a review of the relevant literature, see Willie Langeland and 
Christina Hartgers, "Child Sexual and Physical Abuse and Alco- 
holism/' Journal of Studies on Alcohol 59 (1998): 336— 348. 

8. For a detailed discussion of the gateway concept, see Robert J. 
MacCoun and Peter Reuter, Drug War Heresies (New York: Cam- 
bridge University Press, 2001), pp. 345— 351* 

9. Heather Hammer, David Finkelhor, and Andrea J. Sedlak, 
"Children Abducted fcy Family Members: National Estimates and 
Characteristics," National Incidence Studies d Missing, Abducted, Run- 
away, and Thrownaway Children (U.S. Office of Juvenile Justice and 
Delinquency Prevention, October 2002). 

10. Everyone can profit from reading Edward Tufte s magnificent 
book The Visual Display cf Quantitative Information (Cheshire, Conn.: 
Graphics Press, 1983). A large, related literature includes Jones, How to 
Lie with Charts; and Howard Wainer, Visual Revelations (New York: 
Springer -Verlag, 1997). 

11. A fine introduction to the principle thattheareas of two-dimen- 
sional figures must be proportional to the values represented appears in 
Huff ,How to Lie with Statistics, pp. 66—73. 

12. Stephanie J. Ventura, T. J. Mathews, and Brady E. Hamilton, 
"Births to Teenagers in the United States, 1940-2000,” National Vital 
Statistics Reports 49, no. 10 (September 25, 2001): 1. 

13. For an example of one such critique, see Bjorn Lomborg, The 
Skeptical Environmentalist (Cambridge: Cambridge University Press, 
2001), pp. 22-23. 


1 . Lina Guzman, Laura Lippman, Kristin Anderson Moore, and 
William O'Hare, "How Children Are Doing: The Mismatch Between 
Public Perception and Statistical Reality," Child Trends Research Brief 
(July 2003). 

2. On the history of this issue, see Harold C. Sox Jr. and Steven 
Woloshin, "How Many Deaths Are Due to Medical Error? Getting the 
Number Right," Effective Clinical Practice 3 (2000): 277-282. 

3. Warren Wolfe, "Reporting Hospital Errors Seen as Good Idea," 
Minneapolis Star Tribune, December 1, 1999, p. 3B. After citing the esti- 
mates of forty-four thousand to ninety-eight thousand annual deaths, 
this article notes that "medical accidents caused or contributed to 26 
deaths in Minnesota hospitals between 1994 and 1997" but does not 
address the gulf between the very large national estimates and that 
state's vastly smaller number of deaths attributed to errors. 

4. Rodney A. Hayward and Timothy R Hofer, "Estimating Hospi- 
tal Deaths Due to Medical Errors," Journal of the American Medical 
Association 286 (July 25, 2001): 418. This study was based on patients at 
veterans' hospitals, who may be older than — or otherwise differ 
from — patients in other hospitals. 

5. For a more detailed discussion of these problems withmeasuring 
trends, see Best, Damned Lies and Statistics , pp. 98—109. 

6. Alfred Blumstein and Joel Wallman, eds., The Crime Drop in 
America (New York: Cambridge University Press, 2000); Andrew Kar- 
men, New Yor\Murder Mystery: The True Story Behind the Crime Crash 
of the iggcs (New York: New York University Press, 2000). 

7. For a critique of the media's coverage of this story, see David 
Murray, Joel Schwartz, and S. Robert Lichter, It Ain't Necessarily So: 
How Media Ma\e and Unmade the Scientific Picture of Reality (Lanham, 
Md.: Rowman & Litdefield, 2001), pp. 49—52. 

8 . This example comes from Gerd Gigerenzer, Calculated Risfe 
(New York: Simon & Schuster, 2002), p. 41. 

9. A clear discussion of this point appears in John Brignell, Sorry f 
Wrong Number! The Abuse of Measurement (Great Britain: Brignell Asso- 
ciates, 2000), pp. 46-51. On the politics of the issue, see Gary Taubes, 
"Epidemiology Faces Its Limits," Science 269 (July 14, 1995): 164-169. 

10. Cass R. Sunstein, "Probability Neglect: Emotions, Worst Cases, 
and Law,” Yale Law Journal 1 12 (2002): 61 — 107. 

11. RoseM. Kreider and Jason M. Fields, "Number, Timing, and 

Duration of Marriages and Divorces, 1996," Bureau of the Census, 
Current Population Reports, P70-80 (Washington, D.C., February 
2002), For a study based on a different sample that reaches similar con- 
clusions, see Matthew D. Bramlett and William D. Mosher, "First 
Marriage Dissolution, Divorce, and Remarriage: United States," 
Advance Data /k/m Vital and Health Statistics 323 (May 31, 2001). 

12. For a more detailed discussion of these issues, see Best, “Social 
Progress and Social Problems." 


1 . The Wilmington News Journal stories (all from wire service 
reports) were Lindsey Tanner, "Study Takes a Rare Look at Bullying’s 
Broad Effects," April 25, 2001, p. Ai; Tammy Webber, "Sexual Solici- 
tation Reported ty 20% of Kids Who Use Web," June 20, 2001, p. A6; 
and Lindsey Tanner, “1 in 5 Girls Abused by a Date, Study Suggests," 
August 1, 2001, p. A5. The corresponding journal articles were Tonja 
R. Nansel et ah, "Bullying Behaviors Among U.S. Youth,” Journal of 
the American Medical Association 285 (April 25, 2001): 2094—2100; Kim- 
berly J. Mitchell, David Finkelhor, and Janis Wolak, "Risk Factors for 
and Impact of Online Sexual Solicitation of Youth," Journal cf the 
American Medical Association 285 (June 20, 2001): 3011— 3014; and Jay 
G. Silverman et ah, "Dating Violence Against Adolescent Girls and 
Associated Substance Abuse, Unhealthy Weight Control, Sexual Risk 
Behavior, Pregnancy, and Suicidal ity,” Journal of the American Medical 
Association 286 (August 1, zooi): 572-579. 

2. On the efforts of major medical journals to gain media coverage 
of their articles, see Ellen Ruppel Shell, "The Hippocratic Wars," N&f 
Yor\ Times Magazine, June28, 1998, pp. 34-38. 

3. Nels Ericson, "Addressing the Problem of Juvenile Bullying," 
Fact Sheet 27, June 2001, U.S. Office of Juvenile Justice and Delin- 
quency Prevention, Washington, D.C. 

4. Nansel et ah, "Bullying Behaviors," p. 2095. 

5. Ibid., pp. 2098, 2100. 

^p^JX ^ uiojj IPX,, Aq sapijjE sdtuij^ m9jy 

OSOip ‘ojduil?XO JOJ ‘03g JJOJ ipi?9p 2ui§Utfip Oip SuTlUOUIUDOp SOpiJJl? 
JO S3TJ9S i(ip2u3p 1? 3}OJM. UO}diq 3Uq J3}JodoJ S9tUl£ M3j\[ ’Z1 

£ /Dimopid^; upping 

uoqj^ 3 X J° X^W 3L [X ?? ^pIAT 33 S 'ioqi?}J3p un U1E Aq ppq uoqo 

‘ooqjo OAijaap uu uaaq si?q jouojoo oqj ‘suoiiDipsi ju( moipnog Aui?ui in 
ispuoissqjojd pnpnu oi qsi?} siq^ uStsse oi jomojs sum ipnog oip ji 
- uoo u i 'sjboA joj uu?j£ojd uopiuisiSaj iprap pozipuoissqjojd ^jq2iq 
u pi?q si?q mujojqu^j 'siinj joao oHu^qa spiq SMoqs spioasj spiujojiye^ 
jo sis^pire oil? j^dos v j^ip son£ji? sopp^ 'ipnog oip in Ajpoqjmu isoui 
9 soj sopmns uqoj qapq i^X poiou 3Q3 ax :ii0?s } SinpraAQj ouq • 1 1 

ysopiiiuj~q 3 U , qmjipjB 3 , Qinoq//:dBq ib poisod wp puoiupp^ 
pue iiooz ‘oi A|n{ ‘ 6 Hii =QiAjoig^qunqXjo:is/£jo’:puj 9 qi?'MMM }u 
pojsod £ /Dimopidq oppmg uoqp (qnqq) oip jo q}Apq 3qx„ f sop?p\[ on?d 
-mo^) 'sopp\[ Aq posn osoip uiojj j^qMomos jqjpp imp suou^aqiss^p 
£ursn oqi poiprapa^j OA^q j '(6661 uouiukq iaupp^ 

‘sojuopy) uonmzuzQ ixQjy zqj tnoqy st(tiy^ U9j j :qmoj^ Sunumg ‘sopp^ 
‘ojdurexo joj ‘oog ’smojqojd ipnoA jnoqi? sodAioojo^s mpoui oSuoj 
-pqo sqooq ssoqM ‘soppq y aqip ^ IStSojotdos qnjOAinnoip Aq sisApeire 
ui? uiojj Mojjoq uoissnasip sup in posn?j sjuiod oip jo jsop^[ "oi 

-£ 6 i-d 

K ‘3piDTTlg ( UOHU9A3JJ pue jOJJUO^) 3SE3STQ JOJ SJOJUO^) ’g'Q '6 

■fcq 'd ‘§66i ( i£ qaj^pq ‘tumncQ juofuvjy £ /su33X qapq 8uouiy 
appmg in ssiq urpdxq l t u^D f UBiqBqao^j o( Kj^pq iyi 

'd k g66i ‘ oz qaj^pq 'uonmnsuo'j viuvjiy ^s^qaojAqg 

spiamg uooj^ qa^|q )? ^uuQ^japq ’[ -y y[ iz £ 'd ^ooo^: ‘zz j^qojDQ ‘ssnzjy 
ijWQ^fioj^ m9jy t /syppg; Suouiy dpj sopmng ?? ^ur^p Q ^P^ U V '§ 

- 961 — £61 : (g 6 6 1 ‘oz ipj^pY) L b uofe}} ^pP g M. fa l l v i 
-jojy pun dppjqjojAi £ /£66i— o§6i ^soi^g pajiup :sqjnoy qa^jq Suouiy 
9 pmng ?? XonuoAQjq pue jojjuo^ ^sbqstq joj sjojuo^j ’g’p ’L 

'(oooz ( uopar 

-sirejj^ :’('|g[ £ qnMsunjq M3|q) S9vpni§ uonvjndo^ ot zpinQ s/diunsuo'j y 
:t(iv9(j 01 quig uicug ^uosjojoq ^^IPAV Koussi osoip jo uoissnosip v 

joj 'upujoj sopinSiqui^ ^j^p gHDN 9 X J° 9 X ^y^ S9 Q '9 

Is Down Sharply," November 21, 2001, p. Ai; "In Cold Numbers: A 
Census of the Sept. 11 Victims," April 19, 2002, p. A14; "Death Toll Is 
Near 3,000, but Some Uncertainty over the Count Remains," Septem- 
ber 1 1, 2002, p. G47; and "Sept. 1 1 Death Toll Declines as Two People 
Are Found Alive," November 3, 2002, p. A17. In addition to those 
killed at the World Trade Center, 184 died in the Pentagon attack and 
40 more in the crash of the fourth plane in Pennsylvania, but it was eas- 
ier to count and identify those other victims. The deaths of the 19 
hijackers were also more easily confirmed. 

13. Nina Bernstein, "Thousands of Orphans? An Urban Myth," 
New Yor ^ Timcc, October 2 6, 2001, p. Bi. 

14. Dean Baker, ed., Getting Prices Right: The Debate over the Con- 
sumer Price Index (Armonk, N.Y.: M. E. Sharpe, 1998); Jolie Solomon, 
"An Economic Speedometer Gets an Overhaul," New York [ Timcc, 
December 23, 2001, sec. 3, p. 4. 

15. For example, see Norimitsu Onishi, "African Numbers, Prob- 
lems, and Number Problems," New Yoi\ Times, August 18,2002, sec. 
4? P* 5* 

16. On the dark figure, see Best, Damned Lies and Statistics , pp. 


1. John M. Berry, "Number Crunchers vs. Recession: Seeking 
Official End, Panel Wrestles with One Stubborn Stat," Washington 
Post, July it, 2003, p. Ei; and Berry, "Recession Ended in November of 
2001, But Panel Still Sees Economy Slumping," Washington Post, July 
18, 2003, p. Ei. 

2. James Fallows, "The Early-Decision Racket," Atlantic Monthly 
288 (September 2001): 37-52. 

3. The formula used in 2002 is explained in Robert J. Morse and 
Samuel M. Flanigan, "The Rankings," U.S. News & World Report, 
America's Best Colleges t 2003 Edition, September 2002, pp. 79-81. 

4. The classic work on these issues remains Harold L. Wilensky, 

Organizational Intelligence: Knowledge and Policy in Government and 
Industry (New York: Basic Books, 1967). 

5. The Soviet system has been analyzed in detail; see, for example, 
Stefan Hedlund, Crisis in Soviet Agriculture (New York: St. Martin's, 

6. Alex Berenson, The Number: How the Drive for Quarterly Earn- 
ing Corrupted Wall Street and Corporate America (New York: Random 
House, 2003). 

7. Mark Fazlollah, Michael Matz, and Craig R. McCoy, "How to 
Cut City's Crime Rate: Don't Report It,” Philadelphia Inquirer, Novem- 
ber 1, 1998, p. Ai. 

8. The Enron scandal promises to generate a huge body of litera- 
ture. F or an early introduction, see Loren F ox, Enron: The Rise and Fall 
(Hoboken, N.J.: Wiley, 2003). 

9. Bureau of the Census, Historical Statistics cf the United States 
(Washington, D.C., 1975), p. 55. 

10. National Center for Education Statistics, Digest cf Education 
Statistics, 200 r, Table 41 1; posted at 

11. Many observers have offered competing interpretations of 
trends in SAT scores. For two different, rather subtle discussions, see 
Scott Menard, "Going Down, Going Up: Explaining the Turnaround 
in SAT Scores,” Youth and Society 20 (1988):3-28; and Charles Murray 
and R. J. Herrnstein, "What's Really Behind the SAT-Score Decline?" 
The Public Interest 106 (1992): 32—56. 

12. A huge body of literature focuses on recent testing-centered 
school reforms. For example, generally favorable interpretations may 
be found in Williamson M. Evers and Herbert J. Walberg, eds., School 
Accountability (Stanford, Calif.: Hoover Institution Press, 2002). For a 
more skeptical analysis, see Laura S. Hamilton, Brian M. Stecher, and 
Stephen P. Klein, Maying Sense cf Tat-Based Accountability in Educa- 
tion (Santa Monica, Calif.: Rand, 2002). On evidence of cheating, see 
Brian A. Jacob and Steven D. Levitt, "To Catch a Cheat," Education 
Next 4 (Winter 2004): 69-75. 

13* Thomas J. Kane, Douglas O. Staiger, and Jeffrey Geppert, 
"Randomly Accountable/' Education Next 2 (Spring 2002): 57-61. 

14. Richard J.Lundman and Robert L. Kaufman, “Driving While 
Black," Criminology 41 (2003): 195-220. 

15. Studies of racial profiling are only beginning to appear. See, for 
example, Robin Shepard Engel, Jennifer M. Cainon, and Thomas J. 
Bernard, "Theory and Racial Profiling," Justice Quarterly 19 (zooz): 
249-273; Albert J. Meehan and Michael C. Ponder, "Race and Place: 
The Ecology of Racial Profiling African American Motorists,” Justice 
Quarterly 19 (zooz): 399-430; Michael R. Smith and Geoffrey P. Alpert, 
"Searching for Direction: Courts, Social Science, and the Adjudication 
of Racial Profiling Claims,” Justice Quarterly 19 (zooz): 673—703; and 
Samuel Walker, "Searching for the Denominator,” Research and 
Policy 3 (2001): 63-95. 

16. Walker, "Searching for the Denominator," 63-95. 


1. On stat wars, see Best, Damned Lies and Statistics , pp. 128-159. 

2. On some of the more dubious recent claims to scientific status, 
see Robert Park, Voodoo Science: The Road from Foolishness to Fraud 
(New York: Oxford University Press, 2000). 

3. For studies of some scientists' efforts to ding to marginal posi- 
tions, see H.M. Collins, "Surviving Closure: Post-Rejection Adapta- 
tion and Plurality in Science, "American Sociological Review 65 (2000): 
824-845; Bart Simon, Undead Science: Science Studies and the Afterlife of 
ColdFusion (New Brunswick, N.J.: Rutgers University Press, zooz). 

4. Dorothy Nelkin, Selling Science: How the Press Covers Science and 
Technology (New York: W. H. Freeman, 1987). 

5. The term came into common usage via Peter W. Huber, Galileo's 
Revenge: Jun\Science in the Courtroom (New York: Basic Books, 1991)* 

6. Marcia Angell, Science on Trial The Clash cf Medical Evidence 
and the Law in the Breast Implant Case (New York: Norton, 1996). 

7. For a conservative manifesto on this theme, see Steven J.Milloy, 

Junf^ Science Judo: Self-Defense Against Health Scares and Scam (Wash- 
ington, D.C.: Cato Institute, 2001). 

8. Sheldon Rampton and John Stauber, Trust Us, We're Experts: 
How Industry Manipulates Science and Gambles with Your Future (New 
York: Tar cher/Putnam, 2001), p. 265 

9. Colin McMullan and JohnEyles, "Risky Business: An Analysis 
of Claims making in the Development of an Ontario Drinking Water 
Objective for Tritium," Social Problems 46 (1999): 294— 31 1. 

10. A number of recent commentaries have addressed the spinning 
process. See, for example, George Pitcher, The Death of Spin (Hoboken, 
N .J.: Wiley, 2003); Lynn Smith, "Putting a Spin on the Truth with Sta- 
tistics and Studies/' Los Angeles Times t Juneb, 200 i ? p. Ei. 

11. D'Vera Cohn, "Married-With-Children Still Fading; Census 
Finds Americans Living Alone in 25% of Households," Washington 
Post, May 15, 200 1, p. Ai. 

1 2. Compare Eric Schlosser, Fast Fwd Nation: The Dar\ Side of the 
All-American Meal (New York: Houghton Mifflin, 200 1), pp. 239-243; 
and Douglas J. Besharov, "Growing Overweight and Obesity in Amer- 
ica: The Potential Role of Federal Nutrition Programs," testimony 
before the U.S. Senate Committee on Agriculture, Nutrition, and 
Forestry (American Enterprise Institute, 2003). 

13. Kathleen Maguire and Ann L. Pastore, eds., Sourcebook of 
Criminal Justice Statistics ; 2000 (Washington, D.C.: Bureau of Justice 
Statistics, 2001), p. 248. 

1 4. The best overview of the competing estimates is Tom W. Smith, 
"The Muslim Population of the United States: The Methodology of 
Estimates/' Public Opinion Quarterly 66 (2002): 404—417. For examples 
of the critiques offered by rivals in this debate, see Bill Broadway, 
"Number of U.S. Muslims Depends on Who's Counting," Washington 
Post, November 26, 2001, p. Ai . 

15. Calvin Goldscheider, "Are American Jews Vanishing Again?" 
Contexts 2 (Winter 2003), pp. 1 8-24. On the 2000 National Jewish Pop- 
ulation Survey, see Daniel J. Wakin, "Survey of U.S. Jews Sees a Dip; 
Others Demur," New YoH [ Times, October 9, 200 2, p. A23. 

1 6* A growing literature provides various interpretations of the 
consequences of welfare reform. For overviews offering a range of 
viewpoints, see Douglas J. Besharov and Peter Germanis, "Welfare 
Reform — Four Years Later," The Public Interest 140 (2000): 17 - 35 ; 
Christopher Jencks, "Liberal Lessons from Welfare Reform," The 
American Prospect (Special Supplement, Summer 2002): A9-A12; and 
Sanford F. Schram and Joe Soss, "Success Stories: Welfare Reform, 
Policy Discourse, and the Politics of Research," Annals of the American 
Academy of Political and Social Science 577 (2001): 49 — 65 . 

17. Sam Dillon, "Report Finds Deep Poverty Is on the Rise," New 
Yor\ Times , April 30 , 2003 , p. A18. 


Abductions, child. See Missing 

Absolutism, 147, 149 
Accidental deaths among African 
American teens, 106-8 
Accountants, 113-14 
Aesthetics and graphs, 44, 46, 48, 

African Americans: accidental 
deathsof, 106-8; and census 
data, 14-15; and poverty, 166- 
67; and racial profiling, 137-42; 
suicides among, 104-9 
Alcohol and pregnancy, 55-57 
Apocalyptic scenarios, 66, 72-74, 

Authoritative numbers, 91— 115 
Averages, 27-30 

Bias, xiv— xv, 173 
Birth rates, 58-60 

Blonds, extinction of, 19-20 
Breast cancer, 77-78 
Breast implants, 152-53 
Bullying, 94, 96- 101 
Bureaucratic measures, 22-23. See 
also Official numbers; Organiza- 
tional numbers 

Calculation, 171-73 
Campus drinking, 17 
Cancer: breast, 77-78; lung, 39-40, 
76, 81-82, 154-55 
Causality, 37-42 
Census data, 13-15, 156 
Chartjunk, 50, 52 
Cherry-picking, 157-58, 167 
Coconuts, deaths due to falling, 19, 

College guides/admissions, 17, 119— 

2 5 

Comparisons, 87-88 

Confusing numbers, 26-62 
Consumer Price Index, 112— 13 
Contemporary legends, 5. See also 
Legendary numbers 
Contentious numbers, 144-69 
Correlation, 37-42 
Cost-benefit analysis, 9-12, 87-88 
Counts, 28 

Crime rates, 2, 71, 129 
Crime waves, 4-5, 7 
Critical thinking, 177 

"Data dredging," 157-58, 167 
Death records, 103, hi 
Deaths: from falling coconuts, 19; 
from measles, 17-18; from med 
ical errors, 67-69; at schools, 
2-3,4; f rom sniper attacks, 

83; from suicide, 104—9 
Definitions, broad, 8,162 
Departmental organization, effects 
on education, 175-78 
Divorce, 84-87 
Drug use, 40-41, 157-58 

Educational testing, 22-23, 12 1, 

Ethnic categories, 14-16 
Evidence, 147-51 
Examples, 3-7, 68-69 
Experiments, 79-80 
Expert witnesses, 152 

Facts, 146-47 
Family abductions, 46-48 
Federal Bureau of Investigation, 2, 
7, 9 , 84, 138 

Forgotten numbers, 17-19 
Funding, for research, 17, 95 

Global warming, 73-74 
Graphs, 42—61 

Homelessness, 5-6 

Hospital errors. See Medical errors 

Human life, value of, 1 1- 12 

Incalculable numbers, 7-13 
Ideology, xi—xii, 102, 145, 154, 156, 

JAMA. SeeJoumalcf the American 
Medical Association 
Jewels, statistics as, xii-xiii 
Jews, estimating population of, 159, 

Journal <f the American Medical 
Association , 94- 100 
Junk science, 150, 152-55 

Kidnappings, child. See Missing 

Legendary numbers, 19-24 
Life expectancy, 18 

Magical numbers, 116-43 
Mammograms, 77—78 
Marijuana, 40— 41, 157—58 
Mass transit, 10- 11 
Mean, 29-30 
Measles, 17-18 

Media, news, xvi, 4-5, 64, 73-74, 
81, 95—96, 101-2, 151-52 

Median, 29-30 
Medical errors, 67-69 
Meta-analysis, 82-83 
Missing children, 7-9, 46-47 
Missingnumbers, 1-25 
Muslims, estimating population of, 

Native Americans, 14—15 
Numbers games, 125-30, 142-43 

Official numbers, 103-13, See also 
Bureaucratic measures 
Organizational numbers, 125- 
30, See also Bureaucratic 

Peer review, 91-92 
Percentages, 30-36; calculated the 
wrong way, 32-35; as measures 
of change, 35-36, i87n6 
Pessimism, 88-89 
Police shootings, 16 
Population growth, 72-73 
Pregnancy and alcohol, 55-57 
Press releases, 95, 100, 102, 152 
Probability, 76-78 
Probability neglect, 83 
Progress, 18 
Psychiatrists, 104 
Publication bias. 82 

Racial categories, 14-16 
Racial profiling, 137-42 
Recession, 116-17 
Recordkeeping, 103-13, 114 
Relativism, 147-49 

Religious affiliation, 13 - 14, See also 
Jews; Muslims 

Replication, 82, 101, 114, 151-52 
Reporting, 8-9; corporate, 128-29 
Risks, 65, 74-87, 152-55; measur- 
ing, 79-83 

Scandals, statistical, 113-14 
Scary numbers, 63-90 
Scholarly journals, 91-93,95-96. 
See also Journal of the American 
Medical Association 
Scholastic Aptitude Test, 12 1, 13 1. 

See also Educational testing 
School ratings, 130-36; effect of 
school size on, 136 
School shootings, 1-5 
Science, fraud in, 93, 96. See also 
Junk science; Scientific research 
Scientific research, 94- 103, 148-51; 

funding for, 17, 95 
Selectivity, 57-60 
Sex, thinking about, 54-55 
Sexual activity of high school stu- 
dents, 53-54 

Smoking, 39-40, 74-75, 76, 81-82, 
: 54-55 

Sniper attack, risk of, 83 

Social class, 134-35 

Social construction, xiii, 146, 168, 


Social problems, 64-67, 117-18; 

size of, 65, 67-69 
Software, consequences of, 42, 46, 
48-50,52,57, 172, 176 
"Spinning," 155-58 
Spuriousness, 38-40, 42, 80 

Standardized testing. See Educa- 
tional testing 

Statistical literacy, 170-82 
Statistical literacy movement, 178- 

Statistics instruction, xiii, 171—73 
Stat wars, 145 
Student loans, 50-52 
Suicides among African American 
teens, 104-9 

Tax cuts, 26-27,30 
Teachers, assessing quality of, 22- 
23,12^ !33-35 

Teenagers: and crime, 2, 63; and 
drug use, 157-58; graduation 
rates of, 130-32; and pregnancy, 
58-60, 63; and sexual activity, 
53-54; and statistics instruction, 
171-78; suicides among, 104-9. 
See also Bullying 

Trade-offs, 12, 87-88, 130, 155, 167 
Trends, 60, 66, 70-72 
Truncating graphs, 44—46, 49 

Uncounted numbers, 13-17 
Unemployment, 15, 165 
Urban legends, 5 
US. News & World Report . See 
College guides/admissions 

Vaccinations, 12, 18 

Water, recommended intake, 21 
Welfare: fraud in, 6, 18506; reform 
of, 163-67 

Widowhood, average age at, 20 
World Trade Center attacks: 

counting deaths from, 109- n; 
orphans caused by, in 

Y2K, 66, 72 

COMPOSITOR: BookMatters, Berkeley 

TEXT : ii/i^Granjon 
DISPLAY : Orator, Granjon 
PRINTER AND BINDER: Thomson-Shore, Inc.